@tangle-network/agent-bench 0.3.5 → 0.3.6

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (182) hide show
  1. package/CHANGELOG.md +6 -0
  2. package/README.md +13 -1
  3. package/dist/adapters.d.ts +15 -0
  4. package/dist/adapters.js +43 -0
  5. package/dist/adapters.js.map +1 -0
  6. package/dist/benchmarks/_harness.d.ts +125 -0
  7. package/dist/benchmarks/_harness.js +33 -0
  8. package/dist/benchmarks/_harness.js.map +1 -0
  9. package/dist/benchmarks/aec-bench.d.ts +27 -0
  10. package/dist/benchmarks/aec-bench.js +8 -0
  11. package/dist/benchmarks/aec-bench.js.map +1 -0
  12. package/dist/benchmarks/agentbench.d.ts +16 -0
  13. package/dist/benchmarks/agentbench.js +10 -0
  14. package/dist/benchmarks/agentbench.js.map +1 -0
  15. package/dist/benchmarks/appworld.d.ts +37 -0
  16. package/dist/benchmarks/appworld.js +14 -0
  17. package/dist/benchmarks/appworld.js.map +1 -0
  18. package/dist/benchmarks/bfcl.d.ts +18 -0
  19. package/dist/benchmarks/bfcl.js +10 -0
  20. package/dist/benchmarks/bfcl.js.map +1 -0
  21. package/dist/benchmarks/cad-design.d.ts +45 -0
  22. package/dist/benchmarks/cad-design.js +7 -0
  23. package/dist/benchmarks/cad-design.js.map +1 -0
  24. package/dist/benchmarks/cadbench.d.ts +19 -0
  25. package/dist/benchmarks/cadbench.js +8 -0
  26. package/dist/benchmarks/cadbench.js.map +1 -0
  27. package/dist/benchmarks/cadgenbench.d.ts +22 -0
  28. package/dist/benchmarks/cadgenbench.js +8 -0
  29. package/dist/benchmarks/cadgenbench.js.map +1 -0
  30. package/dist/benchmarks/commit0.d.ts +31 -0
  31. package/dist/benchmarks/commit0.js +10 -0
  32. package/dist/benchmarks/commit0.js.map +1 -0
  33. package/dist/benchmarks/crag.d.ts +14 -0
  34. package/dist/benchmarks/crag.js +9 -0
  35. package/dist/benchmarks/crag.js.map +1 -0
  36. package/dist/benchmarks/dabstep.d.ts +18 -0
  37. package/dist/benchmarks/dabstep.js +10 -0
  38. package/dist/benchmarks/dabstep.js.map +1 -0
  39. package/dist/benchmarks/enterpriseops-gym.d.ts +38 -0
  40. package/dist/benchmarks/enterpriseops-gym.js +10 -0
  41. package/dist/benchmarks/enterpriseops-gym.js.map +1 -0
  42. package/dist/benchmarks/finresearchbench.d.ts +15 -0
  43. package/dist/benchmarks/finresearchbench.js +8 -0
  44. package/dist/benchmarks/finresearchbench.js.map +1 -0
  45. package/dist/benchmarks/finsearchcomp.d.ts +49 -0
  46. package/dist/benchmarks/finsearchcomp.js +7 -0
  47. package/dist/benchmarks/finsearchcomp.js.map +1 -0
  48. package/dist/benchmarks/frames.d.ts +59 -0
  49. package/dist/benchmarks/frames.js +13 -0
  50. package/dist/benchmarks/frames.js.map +1 -0
  51. package/dist/benchmarks/hotpotqa.d.ts +48 -0
  52. package/dist/benchmarks/hotpotqa.js +15 -0
  53. package/dist/benchmarks/hotpotqa.js.map +1 -0
  54. package/dist/benchmarks/humaneval.d.ts +53 -0
  55. package/dist/benchmarks/humaneval.js +15 -0
  56. package/dist/benchmarks/humaneval.js.map +1 -0
  57. package/dist/benchmarks/mind2web.d.ts +41 -0
  58. package/dist/benchmarks/mind2web.js +9 -0
  59. package/dist/benchmarks/mind2web.js.map +1 -0
  60. package/dist/benchmarks/nomiracl.d.ts +15 -0
  61. package/dist/benchmarks/nomiracl.js +9 -0
  62. package/dist/benchmarks/nomiracl.js.map +1 -0
  63. package/dist/benchmarks/open-rag-bench.d.ts +14 -0
  64. package/dist/benchmarks/open-rag-bench.js +9 -0
  65. package/dist/benchmarks/open-rag-bench.js.map +1 -0
  66. package/dist/benchmarks/programbench.d.ts +38 -0
  67. package/dist/benchmarks/programbench.js +10 -0
  68. package/dist/benchmarks/programbench.js.map +1 -0
  69. package/dist/benchmarks/rag-shared.d.ts +42 -0
  70. package/dist/benchmarks/rag-shared.js +39 -0
  71. package/dist/benchmarks/rag-shared.js.map +1 -0
  72. package/dist/benchmarks/ragbench.d.ts +16 -0
  73. package/dist/benchmarks/ragbench.js +9 -0
  74. package/dist/benchmarks/ragbench.js.map +1 -0
  75. package/dist/benchmarks/simpleqa.d.ts +64 -0
  76. package/dist/benchmarks/simpleqa.js +11 -0
  77. package/dist/benchmarks/simpleqa.js.map +1 -0
  78. package/dist/benchmarks/swe-bench.d.ts +56 -0
  79. package/dist/benchmarks/swe-bench.js +14 -0
  80. package/dist/benchmarks/swe-bench.js.map +1 -0
  81. package/dist/benchmarks/t2-ragbench.d.ts +14 -0
  82. package/dist/benchmarks/t2-ragbench.js +9 -0
  83. package/dist/benchmarks/t2-ragbench.js.map +1 -0
  84. package/dist/benchmarks/tau-bench-shared.d.ts +26 -0
  85. package/dist/benchmarks/tau-bench-shared.js +10 -0
  86. package/dist/benchmarks/tau-bench-shared.js.map +1 -0
  87. package/dist/benchmarks/tau2-bench.d.ts +7 -0
  88. package/dist/benchmarks/tau2-bench.js +11 -0
  89. package/dist/benchmarks/tau2-bench.js.map +1 -0
  90. package/dist/benchmarks/tau3-banking.d.ts +15 -0
  91. package/dist/benchmarks/tau3-banking.js +9 -0
  92. package/dist/benchmarks/tau3-banking.js.map +1 -0
  93. package/dist/benchmarks/terminal-bench.d.ts +24 -0
  94. package/dist/benchmarks/terminal-bench.js +8 -0
  95. package/dist/benchmarks/terminal-bench.js.map +1 -0
  96. package/dist/benchmarks/toollm.d.ts +16 -0
  97. package/dist/benchmarks/toollm.js +10 -0
  98. package/dist/benchmarks/toollm.js.map +1 -0
  99. package/dist/benchmarks/trata-hedge.d.ts +32 -0
  100. package/dist/benchmarks/trata-hedge.js +7 -0
  101. package/dist/benchmarks/trata-hedge.js.map +1 -0
  102. package/dist/benchmarks/types.d.ts +107 -0
  103. package/dist/benchmarks/types.js +1 -0
  104. package/dist/benchmarks/types.js.map +1 -0
  105. package/dist/benchmarks/webarena-verified.d.ts +16 -0
  106. package/dist/benchmarks/webarena-verified.js +10 -0
  107. package/dist/benchmarks/webarena-verified.js.map +1 -0
  108. package/dist/chunk-2PVVP7GN.js +197 -0
  109. package/dist/chunk-2PVVP7GN.js.map +1 -0
  110. package/dist/chunk-2XU6OGEN.js +170 -0
  111. package/dist/chunk-2XU6OGEN.js.map +1 -0
  112. package/dist/chunk-53UPUNBZ.js +325 -0
  113. package/dist/chunk-53UPUNBZ.js.map +1 -0
  114. package/dist/chunk-5SBJCB6W.js +144 -0
  115. package/dist/chunk-5SBJCB6W.js.map +1 -0
  116. package/dist/chunk-7WSD27QQ.js +118 -0
  117. package/dist/chunk-7WSD27QQ.js.map +1 -0
  118. package/dist/chunk-C7T7WEK2.js +103 -0
  119. package/dist/chunk-C7T7WEK2.js.map +1 -0
  120. package/dist/chunk-CKUVRZ2T.js +251 -0
  121. package/dist/chunk-CKUVRZ2T.js.map +1 -0
  122. package/dist/chunk-HBSWHQNJ.js +30 -0
  123. package/dist/chunk-HBSWHQNJ.js.map +1 -0
  124. package/dist/chunk-HHXFIHXC.js +116 -0
  125. package/dist/chunk-HHXFIHXC.js.map +1 -0
  126. package/dist/chunk-IFAV6KEM.js +276 -0
  127. package/dist/chunk-IFAV6KEM.js.map +1 -0
  128. package/dist/chunk-INNOYXCP.js +387 -0
  129. package/dist/chunk-INNOYXCP.js.map +1 -0
  130. package/dist/chunk-J3KDJNX2.js +182 -0
  131. package/dist/chunk-J3KDJNX2.js.map +1 -0
  132. package/dist/chunk-JRWWGMK7.js +148 -0
  133. package/dist/chunk-JRWWGMK7.js.map +1 -0
  134. package/dist/chunk-JTHWEDEW.js +32 -0
  135. package/dist/chunk-JTHWEDEW.js.map +1 -0
  136. package/dist/chunk-KDIKRJGB.js +120 -0
  137. package/dist/chunk-KDIKRJGB.js.map +1 -0
  138. package/dist/chunk-LRRD7NAG.js +301 -0
  139. package/dist/chunk-LRRD7NAG.js.map +1 -0
  140. package/dist/chunk-ODT47UAY.js +221 -0
  141. package/dist/chunk-ODT47UAY.js.map +1 -0
  142. package/dist/chunk-PA2ZKHJC.js +230 -0
  143. package/dist/chunk-PA2ZKHJC.js.map +1 -0
  144. package/dist/chunk-PPYSEKFM.js +182 -0
  145. package/dist/chunk-PPYSEKFM.js.map +1 -0
  146. package/dist/chunk-PUIRNYI7.js +189 -0
  147. package/dist/chunk-PUIRNYI7.js.map +1 -0
  148. package/dist/chunk-R36V2VP7.js +169 -0
  149. package/dist/chunk-R36V2VP7.js.map +1 -0
  150. package/dist/chunk-R67DFVLO.js +142 -0
  151. package/dist/chunk-R67DFVLO.js.map +1 -0
  152. package/dist/chunk-SEVJPLZC.js +260 -0
  153. package/dist/chunk-SEVJPLZC.js.map +1 -0
  154. package/dist/chunk-SYDW647C.js +318 -0
  155. package/dist/chunk-SYDW647C.js.map +1 -0
  156. package/dist/chunk-TBKU5XQI.js +228 -0
  157. package/dist/chunk-TBKU5XQI.js.map +1 -0
  158. package/dist/chunk-TSWPNOYM.js +147 -0
  159. package/dist/chunk-TSWPNOYM.js.map +1 -0
  160. package/dist/chunk-UAIOHCUK.js +27 -0
  161. package/dist/chunk-UAIOHCUK.js.map +1 -0
  162. package/dist/chunk-UPAMRDX4.js +233 -0
  163. package/dist/chunk-UPAMRDX4.js.map +1 -0
  164. package/dist/chunk-VQRS7VUC.js +342 -0
  165. package/dist/chunk-VQRS7VUC.js.map +1 -0
  166. package/dist/chunk-X3BTXCJ4.js +262 -0
  167. package/dist/chunk-X3BTXCJ4.js.map +1 -0
  168. package/dist/chunk-X5YKXC6V.js +211 -0
  169. package/dist/chunk-X5YKXC6V.js.map +1 -0
  170. package/dist/chunk-Y6O2OCUO.js +130 -0
  171. package/dist/chunk-Y6O2OCUO.js.map +1 -0
  172. package/dist/chunk-YCGY7UIZ.js +208 -0
  173. package/dist/chunk-YCGY7UIZ.js.map +1 -0
  174. package/dist/chunk-Z7ML6L77.js +162 -0
  175. package/dist/chunk-Z7ML6L77.js.map +1 -0
  176. package/dist/chunk-ZEWMTR5M.js +136 -0
  177. package/dist/chunk-ZEWMTR5M.js.map +1 -0
  178. package/dist/index.d.ts +355 -0
  179. package/dist/index.js +1908 -0
  180. package/dist/index.js.map +1 -0
  181. package/package.json +22 -6
  182. package/scripts/verify-packed-consumer.mjs +12 -1
@@ -0,0 +1,325 @@
1
+ // src/benchmarks/cad-design.ts
2
+ import { execFile } from "child_process";
3
+ import { mkdtemp, readFile, writeFile } from "fs/promises";
4
+ import { tmpdir } from "os";
5
+ import { join } from "path";
6
+ import { promisify } from "util";
7
+ var execFileAsync = promisify(execFile);
8
+ async function openscad(args, cwd) {
9
+ const { stdout, stderr } = await execFileAsync("xvfb-run", ["-a", "openscad", ...args], {
10
+ cwd,
11
+ maxBuffer: 1024 * 1024 * 64,
12
+ timeout: 12e4
13
+ });
14
+ return { stdout, stderr };
15
+ }
16
+ function measureStl(stl, volumesFromStderr) {
17
+ const verts = [];
18
+ const tris = [];
19
+ const re = /vertex\s+(-?[\d.eE+]+)\s+(-?[\d.eE+]+)\s+(-?[\d.eE+]+)/g;
20
+ let m;
21
+ const flat = [];
22
+ while (m = re.exec(stl)) flat.push([Number(m[1]), Number(m[2]), Number(m[3])]);
23
+ for (let i = 0; i + 2 < flat.length; i += 3) tris.push([flat[i], flat[i + 1], flat[i + 2]]);
24
+ for (const v of flat) verts.push(v);
25
+ const xs = verts.map((v) => v[0]);
26
+ const ys = verts.map((v) => v[1]);
27
+ const zs = verts.map((v) => v[2]);
28
+ const min = (a) => Math.min(...a);
29
+ const max = (a) => Math.max(...a);
30
+ const minZ = min(zs);
31
+ const maxZ = max(zs);
32
+ const h = maxZ - minZ || 1;
33
+ const footprint = (lo, hi) => {
34
+ const band = verts.filter((v) => v[2] >= minZ + lo * h && v[2] <= minZ + hi * h);
35
+ if (band.length < 3) return 0;
36
+ const bx = band.map((v) => v[0]);
37
+ const by = band.map((v) => v[1]);
38
+ return (max(bx) - min(bx)) * (max(by) - min(by));
39
+ };
40
+ let vol6 = 0;
41
+ for (const [a, b, c] of tris) {
42
+ vol6 += a[0] * (b[1] * c[2] - b[2] * c[1]) - a[1] * (b[0] * c[2] - b[2] * c[0]) + a[2] * (b[0] * c[1] - b[1] * c[0]);
43
+ }
44
+ const bbox = { x: max(xs) - min(xs), y: max(ys) - min(ys), z: maxZ - minZ };
45
+ return {
46
+ triangles: tris.length,
47
+ volumes: volumesFromStderr,
48
+ bbox,
49
+ baseFootprint: footprint(0, 0.6),
50
+ topFootprint: footprint(0.75, 1),
51
+ solidVolume: Math.abs(vol6) / 6,
52
+ bboxVolume: bbox.x * bbox.y * bbox.z || 1
53
+ };
54
+ }
55
+ function inRange(v, [lo, hi]) {
56
+ return v >= lo && v <= hi;
57
+ }
58
+ function scoreGeometry(g, spec) {
59
+ const checks = {};
60
+ if (spec.volumes) checks.volumes = inRange(g.volumes, spec.volumes);
61
+ if (spec.minTriangles != null) checks.detail = g.triangles >= spec.minTriangles;
62
+ if (spec.bbox?.x) checks.bboxX = inRange(g.bbox.x, spec.bbox.x);
63
+ if (spec.bbox?.y) checks.bboxY = inRange(g.bbox.y, spec.bbox.y);
64
+ if (spec.bbox?.z) checks.bboxZ = inRange(g.bbox.z, spec.bbox.z);
65
+ if (spec.pitchedRoof != null)
66
+ checks.pitchedRoof = g.baseFootprint > 0 && g.topFootprint < spec.pitchedRoof * g.baseFootprint;
67
+ if (spec.hollowBelow != null) checks.hollow = g.solidVolume < spec.hollowBelow * g.bboxVolume;
68
+ const vals = Object.values(checks);
69
+ const score = vals.length ? vals.filter(Boolean).length / vals.length : 0;
70
+ return { checks, score };
71
+ }
72
+ function buildingGold(w, d, sh, ns, rh, t = 3) {
73
+ return `
74
+ w=${w}; d=${d}; sh=${sh}; ns=${ns}; rh=${rh}; t=${t};
75
+ module shell(w,d,h,t){ difference(){ cube([w,d,h]); translate([t,t,-1]) cube([w-2*t,d-2*t,h+2]); } }
76
+ module roof(w,d,h){ rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }
77
+ union(){
78
+ for(s=[0:ns-1]) translate([0,0,s*sh]) shell(w,d,sh,t);
79
+ for(s=[0:ns]) translate([0,0,s*sh-0.5]) cube([w,d,1]);
80
+ translate([0,0,ns*sh]) roof(w,d,rh);
81
+ for(s=[0:ns-1]) for(i=[0:floor((w-28)/22)]) translate([14+i*22,-1,s*sh+10]) cube([10,t+2,12]);
82
+ translate([w/2-7,-1,2]) cube([14,t+2,20]);
83
+ }
84
+ `.trim();
85
+ }
86
+ function buildingTask(o) {
87
+ const { id, desc, w, d, sh, ns, rh } = o;
88
+ const totalWall = ns * sh;
89
+ const totalH = totalWall + rh;
90
+ const r = Math.round;
91
+ const spec = {
92
+ volumes: [1, ns * 4 + 6],
93
+ bbox: {
94
+ x: [r(w * 0.82), r(w * 1.18)],
95
+ y: [r(d * 0.82), r((d + 1) * 1.2)],
96
+ z: [r(totalH * 0.82), r((totalH + 1) * 1.2)]
97
+ },
98
+ minTriangles: 50,
99
+ pitchedRoof: 0.55,
100
+ hollowBelow: 0.72
101
+ };
102
+ const prompt = [
103
+ `Write OpenSCAD source for ${desc}.`,
104
+ "Requirements:",
105
+ `- footprint roughly ${w} (X) by ${d} (Y) units`,
106
+ ns > 1 ? `- ${ns} stories, total wall height ~${totalWall} units` : `- one story, wall height ~${sh} units`,
107
+ "- a PITCHED / gabled roof on top that tapers toward a ridge (NOT flat)",
108
+ "- hollow shell (walls with an interior cavity), not a solid block",
109
+ "- a door opening and several windows",
110
+ "Output ONLY the .scad source, no prose, no code fences."
111
+ ].join("\n");
112
+ return { id, prompt, meta: { spec, gold: buildingGold(w, d, sh, ns, rh) } };
113
+ }
114
+ var BUILDING_CONFIGS = [
115
+ { id: "cottage", desc: "a single-story cottage", w: 50, d: 40, sh: 30, ns: 1, rh: 18 },
116
+ { id: "cabin", desc: "a small one-room cabin", w: 44, d: 36, sh: 26, ns: 1, rh: 22 },
117
+ { id: "bungalow", desc: "a wide single-story bungalow", w: 70, d: 52, sh: 28, ns: 1, rh: 20 },
118
+ { id: "barn", desc: "a long deep barn", w: 60, d: 96, sh: 42, ns: 1, rh: 26 },
119
+ { id: "warehouse", desc: "a large warehouse", w: 110, d: 80, sh: 40, ns: 1, rh: 26 },
120
+ { id: "chapel", desc: "a narrow tall chapel with a steep roof", w: 40, d: 72, sh: 48, ns: 1, rh: 34 },
121
+ { id: "farmhouse", desc: "a two-story farmhouse", w: 72, d: 56, sh: 30, ns: 2, rh: 26 },
122
+ { id: "two-story-manor", desc: "a wide two-story manor", w: 96, d: 64, sh: 32, ns: 2, rh: 28 },
123
+ { id: "townhouse", desc: "a narrow three-story townhouse", w: 42, d: 52, sh: 30, ns: 3, rh: 34 },
124
+ { id: "watchtower", desc: "a tall narrow three-story watchtower", w: 34, d: 34, sh: 30, ns: 3, rh: 32 }
125
+ ];
126
+ var WATER_TOWER_GOLD = `
127
+ $fn=48; r=24; bh=72; t=3; ch=26;
128
+ union(){
129
+ difference(){ cylinder(h=bh, r=r); translate([0,0,t]) cylinder(h=bh, r=r-t); }
130
+ translate([0,0,bh]) cylinder(h=ch, r1=r, r2=0);
131
+ }
132
+ `.trim();
133
+ var AFRAME_GOLD = `
134
+ w=44; d=60; h=66; t=3;
135
+ module tri(w,d,h){ rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }
136
+ union(){
137
+ difference(){ tri(w,d,h); translate([t,0,0]) tri(w-2*t,d,h-2*t*h/d); }
138
+ translate([w/2-7,-1,0]) cube([14,t+2,24]);
139
+ }
140
+ `.trim();
141
+ var TASKS = [
142
+ {
143
+ id: "two-story-house",
144
+ prompt: [
145
+ "Write OpenSCAD source for a two-story house.",
146
+ "Requirements:",
147
+ "- footprint roughly 80 (X) by 60 (Y) units",
148
+ "- two stories, total wall height ~68 units",
149
+ "- a PITCHED / gabled roof on top (NOT a flat roof) \u2014 the roof must taper toward a ridge",
150
+ "- hollow shell (walls with an interior cavity), not a solid block",
151
+ "- at least one door opening and several windows",
152
+ "Output ONLY the .scad source, no prose, no code fences."
153
+ ].join("\n"),
154
+ meta: {
155
+ spec: {
156
+ volumes: [1, 12],
157
+ bbox: { x: [70, 110], y: [50, 80], z: [80, 130] },
158
+ minTriangles: 60,
159
+ pitchedRoof: 0.55,
160
+ hollowBelow: 0.7
161
+ },
162
+ gold: `
163
+ w=80; d=60; sh=34; t=3; rh=28;
164
+ module shell(w,d,h,t){ difference(){ cube([w,d,h]); translate([t,t,-1]) cube([w-2*t,d-2*t,h+2]); } }
165
+ module roof(w,d,h){ translate([0,0,0]) rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }
166
+ union(){
167
+ for(s=[0,1]) translate([0,0,s*sh]) shell(w,d,sh,t);
168
+ for(s=[0,1,2]) translate([0,0,s*sh-1]) cube([w,d,1]);
169
+ translate([0,0,2*sh]) roof(w,d,rh);
170
+ for(s=[0,1]) for(i=[0:2]) translate([16+i*22,-1,s*sh+10]) cube([12,t+2,14]);
171
+ translate([w/2-8,-1,2]) cube([16,t+2,22]);
172
+ }
173
+ `.trim()
174
+ }
175
+ },
176
+ ...BUILDING_CONFIGS.map(buildingTask),
177
+ {
178
+ id: "water-tower",
179
+ prompt: [
180
+ "Write OpenSCAD source for a water tower: a tall hollow cylindrical tank with a conical roof.",
181
+ "Requirements:",
182
+ "- a hollow cylindrical tank, ~48 units in diameter, ~72 units tall, with an interior cavity",
183
+ "- a CONICAL roof on top that tapers to a point (a pitched roof)",
184
+ "- wall thickness around 3 units",
185
+ "Output ONLY the .scad source, no prose, no code fences."
186
+ ].join("\n"),
187
+ meta: {
188
+ spec: {
189
+ volumes: [1, 4],
190
+ bbox: { x: [40, 56], y: [40, 56], z: [88, 110] },
191
+ minTriangles: 60,
192
+ pitchedRoof: 0.55,
193
+ hollowBelow: 0.8
194
+ },
195
+ gold: WATER_TOWER_GOLD
196
+ }
197
+ },
198
+ {
199
+ id: "a-frame-cabin",
200
+ prompt: [
201
+ "Write OpenSCAD source for an A-frame cabin: a steep triangular shell where the roof forms the walls.",
202
+ "Requirements:",
203
+ "- footprint roughly 44 (X) by 60 (Y) units",
204
+ "- a STEEP triangular cross-section ~66 units tall that tapers to a ridge at the top",
205
+ "- hollow inside (an interior cavity), not a solid wedge",
206
+ "- a door opening at the front",
207
+ "Output ONLY the .scad source, no prose, no code fences."
208
+ ].join("\n"),
209
+ meta: {
210
+ spec: {
211
+ volumes: [1, 4],
212
+ bbox: { x: [36, 52], y: [50, 74], z: [56, 80] },
213
+ minTriangles: 30,
214
+ pitchedRoof: 0.45,
215
+ hollowBelow: 0.8
216
+ },
217
+ gold: AFRAME_GOLD
218
+ }
219
+ },
220
+ {
221
+ id: "hex-planter",
222
+ prompt: [
223
+ "Write OpenSCAD source for a hexagonal planter pot.",
224
+ "Requirements:",
225
+ "- a hollow hexagonal prism (6-sided), outer width ~50 units across, height ~45",
226
+ "- open at the top, closed at the bottom (a cavity for soil)",
227
+ "- a wall thickness around 3 units",
228
+ "Output ONLY the .scad source, no prose, no code fences."
229
+ ].join("\n"),
230
+ meta: {
231
+ spec: {
232
+ volumes: [1, 3],
233
+ bbox: { x: [40, 60], y: [40, 60], z: [40, 55] },
234
+ minTriangles: 30,
235
+ hollowBelow: 0.7
236
+ },
237
+ gold: `
238
+ $fn=6; or=25; h=45; t=3;
239
+ difference(){
240
+ cylinder(h=h, r=or);
241
+ translate([0,0,t]) cylinder(h=h, r=or-t);
242
+ }
243
+ `.trim()
244
+ }
245
+ }
246
+ ];
247
+ function createCadDesignAdapter() {
248
+ return {
249
+ name: "cad-design",
250
+ async preflight() {
251
+ try {
252
+ await execFileAsync("xvfb-run", ["-a", "openscad", "--version"], { timeout: 3e4 });
253
+ } catch (err) {
254
+ const msg = err instanceof Error ? err.message : String(err);
255
+ throw new Error(
256
+ `cad-design preflight failed: ${msg}
257
+ Fix: install OpenSCAD + Xvfb (Debian/Ubuntu: sudo apt-get install -y openscad xvfb). The judge runs \`xvfb-run -a openscad -o out.stl model.scad\` \u2014 both must be on PATH.`
258
+ );
259
+ }
260
+ },
261
+ async loadTasks(opts = {}) {
262
+ let tasks = TASKS;
263
+ if (opts.ids) tasks = tasks.filter((t) => opts.ids.includes(t.id));
264
+ if (opts.limit != null) tasks = tasks.slice(0, opts.limit);
265
+ return tasks.map((t) => ({ id: t.id, prompt: t.prompt, metadata: t.meta }));
266
+ },
267
+ async goldArtifact(task) {
268
+ const meta = task.metadata;
269
+ return meta?.gold;
270
+ },
271
+ async judge(task, artifact) {
272
+ const spec = task.metadata.spec;
273
+ const src = artifact.trim();
274
+ if (!src) return { resolved: false, score: 0, detail: "empty artifact" };
275
+ const dir = await mkdtemp(join(tmpdir(), "cad-"));
276
+ const scad = join(dir, "model.scad");
277
+ const stlPath = join(dir, "model.stl");
278
+ const pngPath = join(dir, "model.png");
279
+ await writeFile(scad, src);
280
+ let stderr = "";
281
+ try {
282
+ const r = await openscad(["-o", stlPath, scad], dir);
283
+ stderr = r.stderr;
284
+ } catch (err) {
285
+ const msg = err instanceof Error ? err.message : String(err);
286
+ return { resolved: false, score: 0, detail: `compile failed: ${msg.slice(0, 300)}` };
287
+ }
288
+ try {
289
+ await openscad(
290
+ ["-o", pngPath, "--imgsize=1100,850", "--camera=40,30,40,55,0,25,260", "--colorscheme=Tomorrow", scad],
291
+ dir
292
+ );
293
+ } catch {
294
+ }
295
+ let stl;
296
+ try {
297
+ stl = await readFile(stlPath, "utf8");
298
+ } catch {
299
+ return { resolved: false, score: 0, detail: "compiled but produced no STL (empty geometry)" };
300
+ }
301
+ const volumes = Number(/Volumes:\s*(\d+)/.exec(stderr)?.[1] ?? "1");
302
+ const geo = measureStl(stl, volumes);
303
+ const { checks, score } = scoreGeometry(geo, spec);
304
+ const resolved = score === 1;
305
+ const detail = JSON.stringify({
306
+ checks,
307
+ geo: {
308
+ triangles: geo.triangles,
309
+ volumes: geo.volumes,
310
+ bbox: { x: +geo.bbox.x.toFixed(1), y: +geo.bbox.y.toFixed(1), z: +geo.bbox.z.toFixed(1) },
311
+ hollowRatio: +(geo.solidVolume / geo.bboxVolume).toFixed(3),
312
+ topVsBaseFootprint: geo.baseFootprint ? +(geo.topFootprint / geo.baseFootprint).toFixed(3) : null
313
+ },
314
+ renderPath: pngPath,
315
+ stlPath
316
+ });
317
+ return { resolved, score, detail };
318
+ }
319
+ };
320
+ }
321
+
322
+ export {
323
+ createCadDesignAdapter
324
+ };
325
+ //# sourceMappingURL=chunk-53UPUNBZ.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"sources":["../src/benchmarks/cad-design.ts"],"sourcesContent":["/**\n * CAD-Design adapter. Worker artifact = OpenSCAD source (`.scad`). Judge = the\n * REAL OpenSCAD engine: compile + export STL, then measure the produced solid\n * geometry against the task's spec. Fully deterministic — the authoritative CAD\n * kernel is the gate, not an LLM and not the agent's own claim.\n *\n * This is verifiable-reward CAD: an agent is far better at *writing* exact\n * parametric code than at clicking a GUI, and the kernel either produces the\n * specified geometry or it doesn't. Every judged artifact also renders a PNG,\n * so a refine loop's attempts become a watchable \"model getting better\" reel\n * (run-capsule consumes the renders).\n *\n * Spec checks are geometric and ungameable:\n * - compiles : `openscad -o out.stl` exits 0 (hard gate)\n * - volumes : disconnected-solid count within [min,max]\n * - bbox : overall X/Y/Z extent within bounds (the thing is the right size)\n * - detail : triangle count ≥ floor (not a degenerate single cube)\n * - pitchedRoof : the top band's XY footprint NARROWS vs the base — a flat\n * box can't fake this; a real gabled/hipped roof tapers\n * - hollow : interior cavity present (walls, not a solid block) — the\n * bbox volume materially exceeds the printed solid volume\n *\n * Requires only `openscad` + `xvfb-run` on PATH (no venv, no Docker, no network).\n */\n\nimport { execFile } from 'node:child_process'\nimport { mkdtemp, readFile, writeFile } from 'node:fs/promises'\nimport { tmpdir } from 'node:os'\nimport { join } from 'node:path'\nimport { promisify } from 'node:util'\nimport type { BenchmarkAdapter, BenchScore, BenchTask, LoadOptions } from './types'\n\nconst execFileAsync = promisify(execFile)\n\n/** Spec assertions a CAD task can require. All are deterministic + geometric. */\nexport interface CadSpec {\n volumes?: [number, number]\n bbox?: { x?: [number, number]; y?: [number, number]; z?: [number, number] }\n minTriangles?: number\n /** Top-band XY footprint must be < this fraction of the base footprint. */\n pitchedRoof?: number\n /** Printed solid volume must be < this fraction of the bbox volume (hollow). */\n hollowBelow?: number\n}\n\ninterface CadTaskMeta {\n spec: CadSpec\n /** A known-good .scad that satisfies the spec — the oracle for verify-judge. */\n gold: string\n}\n\n/** Run openscad with xvfb (it needs a GL context even headless). */\nasync function openscad(args: string[], cwd: string): Promise<{ stdout: string; stderr: string }> {\n // -a => auto-pick a free display; openscad writes Facets/Volumes to stderr.\n const { stdout, stderr } = await execFileAsync('xvfb-run', ['-a', 'openscad', ...args], {\n cwd,\n maxBuffer: 1024 * 1024 * 64,\n timeout: 120_000,\n })\n return { stdout, stderr }\n}\n\ninterface Geometry {\n triangles: number\n volumes: number\n bbox: { x: number; y: number; z: number }\n /** XY footprint area (extent_x * extent_y) of the bottom 60% and top 25% z-bands. */\n baseFootprint: number\n topFootprint: number\n /** Convex-ish solid volume estimate from the STL (sum of signed tetra volumes). */\n solidVolume: number\n bboxVolume: number\n}\n\n/** Parse an ASCII STL into the geometric measures the spec checks. Pure. */\nfunction measureStl(stl: string, volumesFromStderr: number): Geometry {\n const verts: Array<[number, number, number]> = []\n const tris: Array<[[number, number, number], [number, number, number], [number, number, number]]> = []\n const re = /vertex\\s+(-?[\\d.eE+]+)\\s+(-?[\\d.eE+]+)\\s+(-?[\\d.eE+]+)/g\n let m: RegExpExecArray | null\n const flat: Array<[number, number, number]> = []\n while ((m = re.exec(stl))) flat.push([Number(m[1]), Number(m[2]), Number(m[3])])\n for (let i = 0; i + 2 < flat.length; i += 3) tris.push([flat[i]!, flat[i + 1]!, flat[i + 2]!])\n for (const v of flat) verts.push(v)\n\n const xs = verts.map((v) => v[0])\n const ys = verts.map((v) => v[1])\n const zs = verts.map((v) => v[2])\n const min = (a: number[]) => Math.min(...a)\n const max = (a: number[]) => Math.max(...a)\n const minZ = min(zs)\n const maxZ = max(zs)\n const h = maxZ - minZ || 1\n\n // Footprint of a z-band = XY extent of the vertices within it.\n const footprint = (lo: number, hi: number): number => {\n const band = verts.filter((v) => v[2] >= minZ + lo * h && v[2] <= minZ + hi * h)\n if (band.length < 3) return 0\n const bx = band.map((v) => v[0])\n const by = band.map((v) => v[1])\n return (max(bx) - min(bx)) * (max(by) - min(by))\n }\n\n // Signed volume via the divergence theorem over triangles (|Σ v0·(v1×v2)/6|).\n let vol6 = 0\n for (const [a, b, c] of tris) {\n vol6 +=\n a[0] * (b[1] * c[2] - b[2] * c[1]) -\n a[1] * (b[0] * c[2] - b[2] * c[0]) +\n a[2] * (b[0] * c[1] - b[1] * c[0])\n }\n const bbox = { x: max(xs) - min(xs), y: max(ys) - min(ys), z: maxZ - minZ }\n return {\n triangles: tris.length,\n volumes: volumesFromStderr,\n bbox,\n baseFootprint: footprint(0, 0.6),\n topFootprint: footprint(0.75, 1),\n solidVolume: Math.abs(vol6) / 6,\n bboxVolume: bbox.x * bbox.y * bbox.z || 1,\n }\n}\n\nfunction inRange(v: number, [lo, hi]: [number, number]): boolean {\n return v >= lo && v <= hi\n}\n\n/** Score the geometry against the spec — each check is a named pass/fail. */\nfunction scoreGeometry(g: Geometry, spec: CadSpec): { checks: Record<string, boolean>; score: number } {\n const checks: Record<string, boolean> = {}\n if (spec.volumes) checks.volumes = inRange(g.volumes, spec.volumes)\n if (spec.minTriangles != null) checks.detail = g.triangles >= spec.minTriangles\n if (spec.bbox?.x) checks.bboxX = inRange(g.bbox.x, spec.bbox.x)\n if (spec.bbox?.y) checks.bboxY = inRange(g.bbox.y, spec.bbox.y)\n if (spec.bbox?.z) checks.bboxZ = inRange(g.bbox.z, spec.bbox.z)\n if (spec.pitchedRoof != null)\n checks.pitchedRoof = g.baseFootprint > 0 && g.topFootprint < spec.pitchedRoof * g.baseFootprint\n if (spec.hollowBelow != null) checks.hollow = g.solidVolume < spec.hollowBelow * g.bboxVolume\n const vals = Object.values(checks)\n const score = vals.length ? vals.filter(Boolean).length / vals.length : 0\n return { checks, score }\n}\n\n/** A parametric multi-story gabled building — the gold generator for the\n * building-family tasks. They share geometry (hollow stacked shells + a gabled\n * roof that tapers to a ridge + cut openings) and differ only in dimensions, so\n * one verified generator yields a spec-passing oracle for each. */\nfunction buildingGold(w: number, d: number, sh: number, ns: number, rh: number, t = 3): string {\n return `\nw=${w}; d=${d}; sh=${sh}; ns=${ns}; rh=${rh}; t=${t};\nmodule shell(w,d,h,t){ difference(){ cube([w,d,h]); translate([t,t,-1]) cube([w-2*t,d-2*t,h+2]); } }\nmodule roof(w,d,h){ rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }\nunion(){\n for(s=[0:ns-1]) translate([0,0,s*sh]) shell(w,d,sh,t);\n for(s=[0:ns]) translate([0,0,s*sh-0.5]) cube([w,d,1]);\n translate([0,0,ns*sh]) roof(w,d,rh);\n for(s=[0:ns-1]) for(i=[0:floor((w-28)/22)]) translate([14+i*22,-1,s*sh+10]) cube([10,t+2,12]);\n translate([w/2-7,-1,2]) cube([14,t+2,20]);\n}\n`.trim()\n}\n\n/** Build a building-family task (prompt + spec + gold) from its dimensions.\n * The spec's bbox is derived from the dims with a ±~18% tolerance so the agent\n * must respect the brief's size; the gold (buildingGold) lands exactly on the\n * nominal dims, so it passes by construction. Keep rh > ns*sh/3 (else the roof\n * base sits inside the top-25% band and pitchedRoof fails). */\nfunction buildingTask(o: { id: string; desc: string; w: number; d: number; sh: number; ns: number; rh: number }): {\n id: string\n prompt: string\n meta: CadTaskMeta\n} {\n const { id, desc, w, d, sh, ns, rh } = o\n const totalWall = ns * sh\n const totalH = totalWall + rh\n const r = Math.round\n const spec: CadSpec = {\n volumes: [1, ns * 4 + 6],\n bbox: {\n x: [r(w * 0.82), r(w * 1.18)],\n y: [r(d * 0.82), r((d + 1) * 1.2)],\n z: [r(totalH * 0.82), r((totalH + 1) * 1.2)],\n },\n minTriangles: 50,\n pitchedRoof: 0.55,\n hollowBelow: 0.72,\n }\n const prompt = [\n `Write OpenSCAD source for ${desc}.`,\n 'Requirements:',\n `- footprint roughly ${w} (X) by ${d} (Y) units`,\n ns > 1 ? `- ${ns} stories, total wall height ~${totalWall} units` : `- one story, wall height ~${sh} units`,\n '- a PITCHED / gabled roof on top that tapers toward a ridge (NOT flat)',\n '- hollow shell (walls with an interior cavity), not a solid block',\n '- a door opening and several windows',\n 'Output ONLY the .scad source, no prose, no code fences.',\n ].join('\\n')\n return { id, prompt, meta: { spec, gold: buildingGold(w, d, sh, ns, rh) } }\n}\n\n/** The building family — one parametric skill (hit the bbox + pitched roof +\n * hollow shell), sampled across footprints, heights, story counts, roof rises. */\nconst BUILDING_CONFIGS: Array<{ id: string; desc: string; w: number; d: number; sh: number; ns: number; rh: number }> = [\n { id: 'cottage', desc: 'a single-story cottage', w: 50, d: 40, sh: 30, ns: 1, rh: 18 },\n { id: 'cabin', desc: 'a small one-room cabin', w: 44, d: 36, sh: 26, ns: 1, rh: 22 },\n { id: 'bungalow', desc: 'a wide single-story bungalow', w: 70, d: 52, sh: 28, ns: 1, rh: 20 },\n { id: 'barn', desc: 'a long deep barn', w: 60, d: 96, sh: 42, ns: 1, rh: 26 },\n { id: 'warehouse', desc: 'a large warehouse', w: 110, d: 80, sh: 40, ns: 1, rh: 26 },\n { id: 'chapel', desc: 'a narrow tall chapel with a steep roof', w: 40, d: 72, sh: 48, ns: 1, rh: 34 },\n { id: 'farmhouse', desc: 'a two-story farmhouse', w: 72, d: 56, sh: 30, ns: 2, rh: 26 },\n { id: 'two-story-manor', desc: 'a wide two-story manor', w: 96, d: 64, sh: 32, ns: 2, rh: 28 },\n { id: 'townhouse', desc: 'a narrow three-story townhouse', w: 42, d: 52, sh: 30, ns: 3, rh: 34 },\n { id: 'watchtower', desc: 'a tall narrow three-story watchtower', w: 34, d: 34, sh: 30, ns: 3, rh: 32 },\n]\n\n/** A hollow cylindrical water tower with a conical roof — a non-box shape that\n * still exercises pitched-roof (the cone tapers to an apex) + hollow + bbox. */\nconst WATER_TOWER_GOLD = `\n$fn=48; r=24; bh=72; t=3; ch=26;\nunion(){\n difference(){ cylinder(h=bh, r=r); translate([0,0,t]) cylinder(h=bh, r=r-t); }\n translate([0,0,bh]) cylinder(h=ch, r1=r, r2=0);\n}\n`.trim()\n\n/** An A-frame: the steep roof IS the walls — a hollow triangular prism. The most\n * extreme taper in the set; tests that the agent can build a non-rectangular shell. */\nconst AFRAME_GOLD = `\nw=44; d=60; h=66; t=3;\nmodule tri(w,d,h){ rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }\nunion(){\n difference(){ tri(w,d,h); translate([t,0,0]) tri(w-2*t,d,h-2*t*h/d); }\n translate([w/2-7,-1,0]) cube([14,t+2,24]);\n}\n`.trim()\n\nconst TASKS: Array<{ id: string; prompt: string; meta: CadTaskMeta }> = [\n {\n id: 'two-story-house',\n prompt: [\n 'Write OpenSCAD source for a two-story house.',\n 'Requirements:',\n '- footprint roughly 80 (X) by 60 (Y) units',\n '- two stories, total wall height ~68 units',\n '- a PITCHED / gabled roof on top (NOT a flat roof) — the roof must taper toward a ridge',\n '- hollow shell (walls with an interior cavity), not a solid block',\n '- at least one door opening and several windows',\n 'Output ONLY the .scad source, no prose, no code fences.',\n ].join('\\n'),\n meta: {\n spec: {\n volumes: [1, 12],\n bbox: { x: [70, 110], y: [50, 80], z: [80, 130] },\n minTriangles: 60,\n pitchedRoof: 0.55,\n hollowBelow: 0.7,\n },\n gold: `\nw=80; d=60; sh=34; t=3; rh=28;\nmodule shell(w,d,h,t){ difference(){ cube([w,d,h]); translate([t,t,-1]) cube([w-2*t,d-2*t,h+2]); } }\nmodule roof(w,d,h){ translate([0,0,0]) rotate([90,0,90]) linear_extrude(height=w) polygon([[0,0],[d,0],[d/2,h]]); }\nunion(){\n for(s=[0,1]) translate([0,0,s*sh]) shell(w,d,sh,t);\n for(s=[0,1,2]) translate([0,0,s*sh-1]) cube([w,d,1]);\n translate([0,0,2*sh]) roof(w,d,rh);\n for(s=[0,1]) for(i=[0:2]) translate([16+i*22,-1,s*sh+10]) cube([12,t+2,14]);\n translate([w/2-8,-1,2]) cube([16,t+2,22]);\n}\n`.trim(),\n },\n },\n ...BUILDING_CONFIGS.map(buildingTask),\n {\n id: 'water-tower',\n prompt: [\n 'Write OpenSCAD source for a water tower: a tall hollow cylindrical tank with a conical roof.',\n 'Requirements:',\n '- a hollow cylindrical tank, ~48 units in diameter, ~72 units tall, with an interior cavity',\n '- a CONICAL roof on top that tapers to a point (a pitched roof)',\n '- wall thickness around 3 units',\n 'Output ONLY the .scad source, no prose, no code fences.',\n ].join('\\n'),\n meta: {\n spec: {\n volumes: [1, 4],\n bbox: { x: [40, 56], y: [40, 56], z: [88, 110] },\n minTriangles: 60,\n pitchedRoof: 0.55,\n hollowBelow: 0.8,\n },\n gold: WATER_TOWER_GOLD,\n },\n },\n {\n id: 'a-frame-cabin',\n prompt: [\n 'Write OpenSCAD source for an A-frame cabin: a steep triangular shell where the roof forms the walls.',\n 'Requirements:',\n '- footprint roughly 44 (X) by 60 (Y) units',\n '- a STEEP triangular cross-section ~66 units tall that tapers to a ridge at the top',\n '- hollow inside (an interior cavity), not a solid wedge',\n '- a door opening at the front',\n 'Output ONLY the .scad source, no prose, no code fences.',\n ].join('\\n'),\n meta: {\n spec: {\n volumes: [1, 4],\n bbox: { x: [36, 52], y: [50, 74], z: [56, 80] },\n minTriangles: 30,\n pitchedRoof: 0.45,\n hollowBelow: 0.8,\n },\n gold: AFRAME_GOLD,\n },\n },\n {\n id: 'hex-planter',\n prompt: [\n 'Write OpenSCAD source for a hexagonal planter pot.',\n 'Requirements:',\n '- a hollow hexagonal prism (6-sided), outer width ~50 units across, height ~45',\n '- open at the top, closed at the bottom (a cavity for soil)',\n '- a wall thickness around 3 units',\n 'Output ONLY the .scad source, no prose, no code fences.',\n ].join('\\n'),\n meta: {\n spec: {\n volumes: [1, 3],\n bbox: { x: [40, 60], y: [40, 60], z: [40, 55] },\n minTriangles: 30,\n hollowBelow: 0.7,\n },\n gold: `\n$fn=6; or=25; h=45; t=3;\ndifference(){\n cylinder(h=h, r=or);\n translate([0,0,t]) cylinder(h=h, r=or-t);\n}\n`.trim(),\n },\n },\n]\n\nexport function createCadDesignAdapter(): BenchmarkAdapter {\n return {\n name: 'cad-design',\n\n async preflight() {\n try {\n await execFileAsync('xvfb-run', ['-a', 'openscad', '--version'], { timeout: 30_000 })\n } catch (err) {\n const msg = err instanceof Error ? err.message : String(err)\n throw new Error(\n `cad-design preflight failed: ${msg}\\n` +\n `Fix: install OpenSCAD + Xvfb (Debian/Ubuntu: sudo apt-get install -y openscad xvfb). ` +\n `The judge runs \\`xvfb-run -a openscad -o out.stl model.scad\\` — both must be on PATH.`,\n )\n }\n },\n\n async loadTasks(opts: LoadOptions = {}) {\n let tasks = TASKS\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 return tasks.map((t): BenchTask => ({ id: t.id, prompt: t.prompt, metadata: t.meta as unknown as Record<string, unknown> }))\n },\n\n async goldArtifact(task: BenchTask) {\n const meta = task.metadata as unknown as CadTaskMeta | undefined\n return meta?.gold\n },\n\n async judge(task: BenchTask, artifact: string): Promise<BenchScore> {\n const spec = (task.metadata as unknown as CadTaskMeta).spec\n const src = artifact.trim()\n if (!src) return { resolved: false, score: 0, detail: 'empty artifact' }\n\n const dir = await mkdtemp(join(tmpdir(), 'cad-'))\n const scad = join(dir, 'model.scad')\n const stlPath = join(dir, 'model.stl')\n const pngPath = join(dir, 'model.png')\n await writeFile(scad, src)\n\n // GATE 1: compile + export STL. Nonzero exit / CGAL error => does not compile.\n let stderr = ''\n try {\n const r = await openscad(['-o', stlPath, scad], dir)\n stderr = r.stderr\n } catch (err) {\n const msg = err instanceof Error ? err.message : String(err)\n return { resolved: false, score: 0, detail: `compile failed: ${msg.slice(0, 300)}` }\n }\n\n // Render a PNG alongside (the artifact run-capsule animates). Non-fatal.\n try {\n await openscad(\n ['-o', pngPath, '--imgsize=1100,850', '--camera=40,30,40,55,0,25,260', '--colorscheme=Tomorrow', scad],\n dir,\n )\n } catch {\n /* render is for the reel; geometry gate already has the STL */\n }\n\n let stl: string\n try {\n stl = await readFile(stlPath, 'utf8')\n } catch {\n return { resolved: false, score: 0, detail: 'compiled but produced no STL (empty geometry)' }\n }\n const volumes = Number(/Volumes:\\s*(\\d+)/.exec(stderr)?.[1] ?? '1')\n const geo = measureStl(stl, volumes)\n const { checks, score } = scoreGeometry(geo, spec)\n const resolved = score === 1\n const detail = JSON.stringify({\n checks,\n geo: {\n triangles: geo.triangles,\n volumes: geo.volumes,\n bbox: { x: +geo.bbox.x.toFixed(1), y: +geo.bbox.y.toFixed(1), z: +geo.bbox.z.toFixed(1) },\n hollowRatio: +(geo.solidVolume / geo.bboxVolume).toFixed(3),\n topVsBaseFootprint: geo.baseFootprint ? +(geo.topFootprint / geo.baseFootprint).toFixed(3) : null,\n },\n renderPath: pngPath,\n stlPath,\n })\n return { resolved, score, detail }\n },\n }\n}\n"],"mappings":";AAyBA,SAAS,gBAAgB;AACzB,SAAS,SAAS,UAAU,iBAAiB;AAC7C,SAAS,cAAc;AACvB,SAAS,YAAY;AACrB,SAAS,iBAAiB;AAG1B,IAAM,gBAAgB,UAAU,QAAQ;AAoBxC,eAAe,SAAS,MAAgB,KAA0D;AAEhG,QAAM,EAAE,QAAQ,OAAO,IAAI,MAAM,cAAc,YAAY,CAAC,MAAM,YAAY,GAAG,IAAI,GAAG;AAAA,IACtF;AAAA,IACA,WAAW,OAAO,OAAO;AAAA,IACzB,SAAS;AAAA,EACX,CAAC;AACD,SAAO,EAAE,QAAQ,OAAO;AAC1B;AAeA,SAAS,WAAW,KAAa,mBAAqC;AACpE,QAAM,QAAyC,CAAC;AAChD,QAAM,OAA8F,CAAC;AACrG,QAAM,KAAK;AACX,MAAI;AACJ,QAAM,OAAwC,CAAC;AAC/C,SAAQ,IAAI,GAAG,KAAK,GAAG,EAAI,MAAK,KAAK,CAAC,OAAO,EAAE,CAAC,CAAC,GAAG,OAAO,EAAE,CAAC,CAAC,GAAG,OAAO,EAAE,CAAC,CAAC,CAAC,CAAC;AAC/E,WAAS,IAAI,GAAG,IAAI,IAAI,KAAK,QAAQ,KAAK,EAAG,MAAK,KAAK,CAAC,KAAK,CAAC,GAAI,KAAK,IAAI,CAAC,GAAI,KAAK,IAAI,CAAC,CAAE,CAAC;AAC7F,aAAW,KAAK,KAAM,OAAM,KAAK,CAAC;AAElC,QAAM,KAAK,MAAM,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;AAChC,QAAM,KAAK,MAAM,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;AAChC,QAAM,KAAK,MAAM,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;AAChC,QAAM,MAAM,CAAC,MAAgB,KAAK,IAAI,GAAG,CAAC;AAC1C,QAAM,MAAM,CAAC,MAAgB,KAAK,IAAI,GAAG,CAAC;AAC1C,QAAM,OAAO,IAAI,EAAE;AACnB,QAAM,OAAO,IAAI,EAAE;AACnB,QAAM,IAAI,OAAO,QAAQ;AAGzB,QAAM,YAAY,CAAC,IAAY,OAAuB;AACpD,UAAM,OAAO,MAAM,OAAO,CAAC,MAAM,EAAE,CAAC,KAAK,OAAO,KAAK,KAAK,EAAE,CAAC,KAAK,OAAO,KAAK,CAAC;AAC/E,QAAI,KAAK,SAAS,EAAG,QAAO;AAC5B,UAAM,KAAK,KAAK,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;AAC/B,UAAM,KAAK,KAAK,IAAI,CAAC,MAAM,EAAE,CAAC,CAAC;AAC/B,YAAQ,IAAI,EAAE,IAAI,IAAI,EAAE,MAAM,IAAI,EAAE,IAAI,IAAI,EAAE;AAAA,EAChD;AAGA,MAAI,OAAO;AACX,aAAW,CAAC,GAAG,GAAG,CAAC,KAAK,MAAM;AAC5B,YACE,EAAE,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,KAChC,EAAE,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,KAChC,EAAE,CAAC,KAAK,EAAE,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC,IAAI,EAAE,CAAC;AAAA,EACpC;AACA,QAAM,OAAO,EAAE,GAAG,IAAI,EAAE,IAAI,IAAI,EAAE,GAAG,GAAG,IAAI,EAAE,IAAI,IAAI,EAAE,GAAG,GAAG,OAAO,KAAK;AAC1E,SAAO;AAAA,IACL,WAAW,KAAK;AAAA,IAChB,SAAS;AAAA,IACT;AAAA,IACA,eAAe,UAAU,GAAG,GAAG;AAAA,IAC/B,cAAc,UAAU,MAAM,CAAC;AAAA,IAC/B,aAAa,KAAK,IAAI,IAAI,IAAI;AAAA,IAC9B,YAAY,KAAK,IAAI,KAAK,IAAI,KAAK,KAAK;AAAA,EAC1C;AACF;AAEA,SAAS,QAAQ,GAAW,CAAC,IAAI,EAAE,GAA8B;AAC/D,SAAO,KAAK,MAAM,KAAK;AACzB;AAGA,SAAS,cAAc,GAAa,MAAmE;AACrG,QAAM,SAAkC,CAAC;AACzC,MAAI,KAAK,QAAS,QAAO,UAAU,QAAQ,EAAE,SAAS,KAAK,OAAO;AAClE,MAAI,KAAK,gBAAgB,KAAM,QAAO,SAAS,EAAE,aAAa,KAAK;AACnE,MAAI,KAAK,MAAM,EAAG,QAAO,QAAQ,QAAQ,EAAE,KAAK,GAAG,KAAK,KAAK,CAAC;AAC9D,MAAI,KAAK,MAAM,EAAG,QAAO,QAAQ,QAAQ,EAAE,KAAK,GAAG,KAAK,KAAK,CAAC;AAC9D,MAAI,KAAK,MAAM,EAAG,QAAO,QAAQ,QAAQ,EAAE,KAAK,GAAG,KAAK,KAAK,CAAC;AAC9D,MAAI,KAAK,eAAe;AACtB,WAAO,cAAc,EAAE,gBAAgB,KAAK,EAAE,eAAe,KAAK,cAAc,EAAE;AACpF,MAAI,KAAK,eAAe,KAAM,QAAO,SAAS,EAAE,cAAc,KAAK,cAAc,EAAE;AACnF,QAAM,OAAO,OAAO,OAAO,MAAM;AACjC,QAAM,QAAQ,KAAK,SAAS,KAAK,OAAO,OAAO,EAAE,SAAS,KAAK,SAAS;AACxE,SAAO,EAAE,QAAQ,MAAM;AACzB;AAMA,SAAS,aAAa,GAAW,GAAW,IAAY,IAAY,IAAY,IAAI,GAAW;AAC7F,SAAO;AAAA,IACL,CAAC,OAAO,CAAC,QAAQ,EAAE,QAAQ,EAAE,QAAQ,EAAE,OAAO,CAAC;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAUjD,KAAK;AACP;AAOA,SAAS,aAAa,GAIpB;AACA,QAAM,EAAE,IAAI,MAAM,GAAG,GAAG,IAAI,IAAI,GAAG,IAAI;AACvC,QAAM,YAAY,KAAK;AACvB,QAAM,SAAS,YAAY;AAC3B,QAAM,IAAI,KAAK;AACf,QAAM,OAAgB;AAAA,IACpB,SAAS,CAAC,GAAG,KAAK,IAAI,CAAC;AAAA,IACvB,MAAM;AAAA,MACJ,GAAG,CAAC,EAAE,IAAI,IAAI,GAAG,EAAE,IAAI,IAAI,CAAC;AAAA,MAC5B,GAAG,CAAC,EAAE,IAAI,IAAI,GAAG,GAAG,IAAI,KAAK,GAAG,CAAC;AAAA,MACjC,GAAG,CAAC,EAAE,SAAS,IAAI,GAAG,GAAG,SAAS,KAAK,GAAG,CAAC;AAAA,IAC7C;AAAA,IACA,cAAc;AAAA,IACd,aAAa;AAAA,IACb,aAAa;AAAA,EACf;AACA,QAAM,SAAS;AAAA,IACb,6BAA6B,IAAI;AAAA,IACjC;AAAA,IACA,uBAAuB,CAAC,WAAW,CAAC;AAAA,IACpC,KAAK,IAAI,KAAK,EAAE,gCAAgC,SAAS,WAAW,6BAA6B,EAAE;AAAA,IACnG;AAAA,IACA;AAAA,IACA;AAAA,IACA;AAAA,EACF,EAAE,KAAK,IAAI;AACX,SAAO,EAAE,IAAI,QAAQ,MAAM,EAAE,MAAM,MAAM,aAAa,GAAG,GAAG,IAAI,IAAI,EAAE,EAAE,EAAE;AAC5E;AAIA,IAAM,mBAAkH;AAAA,EACtH,EAAE,IAAI,WAAW,MAAM,0BAA0B,GAAG,IAAI,GAAG,IAAI,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG;AAAA,EACrF,EAAE,IAAI,SAAS,MAAM,0BAA0B,GAAG,IAAI,GAAG,IAAI,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG;AAAA,EACnF,EAAE,IAAI,YAAY,MAAM,gCAAgC,GAAG,IAAI,GAAG,IAAI,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG;AAAA,EAC5F,EAAE,IAAI,QAAQ,MAAM,oBAAoB,GAAG,IAAI,GAAG,IAAI,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG;AAAA,EAC5E,EAAE,IAAI,aAAa,MAAM,qBAAqB,GAAG,KAAK,GAAG,IAAI,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG;AAAA,EACnF,EAAE,IAAI,UAAU,MAAM,0CAA0C,GAAG,IAAI,GAAG,IAAI,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG;AAAA,EACpG,EAAE,IAAI,aAAa,MAAM,yBAAyB,GAAG,IAAI,GAAG,IAAI,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG;AAAA,EACtF,EAAE,IAAI,mBAAmB,MAAM,0BAA0B,GAAG,IAAI,GAAG,IAAI,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG;AAAA,EAC7F,EAAE,IAAI,aAAa,MAAM,kCAAkC,GAAG,IAAI,GAAG,IAAI,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG;AAAA,EAC/F,EAAE,IAAI,cAAc,MAAM,wCAAwC,GAAG,IAAI,GAAG,IAAI,IAAI,IAAI,IAAI,GAAG,IAAI,GAAG;AACxG;AAIA,IAAM,mBAAmB;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMvB,KAAK;AAIP,IAAM,cAAc;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAOlB,KAAK;AAEP,IAAM,QAAkE;AAAA,EACtE;AAAA,IACE,IAAI;AAAA,IACJ,QAAQ;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,IAAI;AAAA,IACX,MAAM;AAAA,MACJ,MAAM;AAAA,QACJ,SAAS,CAAC,GAAG,EAAE;AAAA,QACf,MAAM,EAAE,GAAG,CAAC,IAAI,GAAG,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,GAAG,CAAC,IAAI,GAAG,EAAE;AAAA,QAChD,cAAc;AAAA,QACd,aAAa;AAAA,QACb,aAAa;AAAA,MACf;AAAA,MACA,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAWV,KAAK;AAAA,IACH;AAAA,EACF;AAAA,EACA,GAAG,iBAAiB,IAAI,YAAY;AAAA,EACpC;AAAA,IACE,IAAI;AAAA,IACJ,QAAQ;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,IAAI;AAAA,IACX,MAAM;AAAA,MACJ,MAAM;AAAA,QACJ,SAAS,CAAC,GAAG,CAAC;AAAA,QACd,MAAM,EAAE,GAAG,CAAC,IAAI,EAAE,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,GAAG,CAAC,IAAI,GAAG,EAAE;AAAA,QAC/C,cAAc;AAAA,QACd,aAAa;AAAA,QACb,aAAa;AAAA,MACf;AAAA,MACA,MAAM;AAAA,IACR;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,QAAQ;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,IAAI;AAAA,IACX,MAAM;AAAA,MACJ,MAAM;AAAA,QACJ,SAAS,CAAC,GAAG,CAAC;AAAA,QACd,MAAM,EAAE,GAAG,CAAC,IAAI,EAAE,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,GAAG,CAAC,IAAI,EAAE,EAAE;AAAA,QAC9C,cAAc;AAAA,QACd,aAAa;AAAA,QACb,aAAa;AAAA,MACf;AAAA,MACA,MAAM;AAAA,IACR;AAAA,EACF;AAAA,EACA;AAAA,IACE,IAAI;AAAA,IACJ,QAAQ;AAAA,MACN;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,MACA;AAAA,IACF,EAAE,KAAK,IAAI;AAAA,IACX,MAAM;AAAA,MACJ,MAAM;AAAA,QACJ,SAAS,CAAC,GAAG,CAAC;AAAA,QACd,MAAM,EAAE,GAAG,CAAC,IAAI,EAAE,GAAG,GAAG,CAAC,IAAI,EAAE,GAAG,GAAG,CAAC,IAAI,EAAE,EAAE;AAAA,QAC9C,cAAc;AAAA,QACd,aAAa;AAAA,MACf;AAAA,MACA,MAAM;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAMV,KAAK;AAAA,IACH;AAAA,EACF;AACF;AAEO,SAAS,yBAA2C;AACzD,SAAO;AAAA,IACL,MAAM;AAAA,IAEN,MAAM,YAAY;AAChB,UAAI;AACF,cAAM,cAAc,YAAY,CAAC,MAAM,YAAY,WAAW,GAAG,EAAE,SAAS,IAAO,CAAC;AAAA,MACtF,SAAS,KAAK;AACZ,cAAM,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC3D,cAAM,IAAI;AAAA,UACR,gCAAgC,GAAG;AAAA;AAAA,QAGrC;AAAA,MACF;AAAA,IACF;AAAA,IAEA,MAAM,UAAU,OAAoB,CAAC,GAAG;AACtC,UAAI,QAAQ;AACZ,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,aAAO,MAAM,IAAI,CAAC,OAAkB,EAAE,IAAI,EAAE,IAAI,QAAQ,EAAE,QAAQ,UAAU,EAAE,KAA2C,EAAE;AAAA,IAC7H;AAAA,IAEA,MAAM,aAAa,MAAiB;AAClC,YAAM,OAAO,KAAK;AAClB,aAAO,MAAM;AAAA,IACf;AAAA,IAEA,MAAM,MAAM,MAAiB,UAAuC;AAClE,YAAM,OAAQ,KAAK,SAAoC;AACvD,YAAM,MAAM,SAAS,KAAK;AAC1B,UAAI,CAAC,IAAK,QAAO,EAAE,UAAU,OAAO,OAAO,GAAG,QAAQ,iBAAiB;AAEvE,YAAM,MAAM,MAAM,QAAQ,KAAK,OAAO,GAAG,MAAM,CAAC;AAChD,YAAM,OAAO,KAAK,KAAK,YAAY;AACnC,YAAM,UAAU,KAAK,KAAK,WAAW;AACrC,YAAM,UAAU,KAAK,KAAK,WAAW;AACrC,YAAM,UAAU,MAAM,GAAG;AAGzB,UAAI,SAAS;AACb,UAAI;AACF,cAAM,IAAI,MAAM,SAAS,CAAC,MAAM,SAAS,IAAI,GAAG,GAAG;AACnD,iBAAS,EAAE;AAAA,MACb,SAAS,KAAK;AACZ,cAAM,MAAM,eAAe,QAAQ,IAAI,UAAU,OAAO,GAAG;AAC3D,eAAO,EAAE,UAAU,OAAO,OAAO,GAAG,QAAQ,mBAAmB,IAAI,MAAM,GAAG,GAAG,CAAC,GAAG;AAAA,MACrF;AAGA,UAAI;AACF,cAAM;AAAA,UACJ,CAAC,MAAM,SAAS,sBAAsB,iCAAiC,0BAA0B,IAAI;AAAA,UACrG;AAAA,QACF;AAAA,MACF,QAAQ;AAAA,MAER;AAEA,UAAI;AACJ,UAAI;AACF,cAAM,MAAM,SAAS,SAAS,MAAM;AAAA,MACtC,QAAQ;AACN,eAAO,EAAE,UAAU,OAAO,OAAO,GAAG,QAAQ,gDAAgD;AAAA,MAC9F;AACA,YAAM,UAAU,OAAO,mBAAmB,KAAK,MAAM,IAAI,CAAC,KAAK,GAAG;AAClE,YAAM,MAAM,WAAW,KAAK,OAAO;AACnC,YAAM,EAAE,QAAQ,MAAM,IAAI,cAAc,KAAK,IAAI;AACjD,YAAM,WAAW,UAAU;AAC3B,YAAM,SAAS,KAAK,UAAU;AAAA,QAC5B;AAAA,QACA,KAAK;AAAA,UACH,WAAW,IAAI;AAAA,UACf,SAAS,IAAI;AAAA,UACb,MAAM,EAAE,GAAG,CAAC,IAAI,KAAK,EAAE,QAAQ,CAAC,GAAG,GAAG,CAAC,IAAI,KAAK,EAAE,QAAQ,CAAC,GAAG,GAAG,CAAC,IAAI,KAAK,EAAE,QAAQ,CAAC,EAAE;AAAA,UACxF,aAAa,EAAE,IAAI,cAAc,IAAI,YAAY,QAAQ,CAAC;AAAA,UAC1D,oBAAoB,IAAI,gBAAgB,EAAE,IAAI,eAAe,IAAI,eAAe,QAAQ,CAAC,IAAI;AAAA,QAC/F;AAAA,QACA,YAAY;AAAA,QACZ;AAAA,MACF,CAAC;AACD,aAAO,EAAE,UAAU,OAAO,OAAO;AAAA,IACnC;AAAA,EACF;AACF;","names":[]}
@@ -0,0 +1,144 @@
1
+ import {
2
+ createToolLlmAdapter
3
+ } from "./chunk-2PVVP7GN.js";
4
+ import {
5
+ createTrataHedgeAdapter
6
+ } from "./chunk-VQRS7VUC.js";
7
+ import {
8
+ createWebArenaVerifiedAdapter
9
+ } from "./chunk-Z7ML6L77.js";
10
+ import {
11
+ createRagBenchAdapter
12
+ } from "./chunk-R67DFVLO.js";
13
+ import {
14
+ createSimpleQaAdapter
15
+ } from "./chunk-SEVJPLZC.js";
16
+ import {
17
+ createSweBenchAdapter
18
+ } from "./chunk-CKUVRZ2T.js";
19
+ import {
20
+ createT2RagBenchAdapter
21
+ } from "./chunk-Y6O2OCUO.js";
22
+ import {
23
+ createTau2BenchAdapter
24
+ } from "./chunk-HBSWHQNJ.js";
25
+ import {
26
+ createTau3BankingAdapter
27
+ } from "./chunk-UAIOHCUK.js";
28
+ import {
29
+ createTerminalBenchAdapter
30
+ } from "./chunk-JRWWGMK7.js";
31
+ import {
32
+ createFramesAdapter
33
+ } from "./chunk-INNOYXCP.js";
34
+ import {
35
+ createHotpotqaAdapter
36
+ } from "./chunk-PA2ZKHJC.js";
37
+ import {
38
+ createHumanEvalAdapter
39
+ } from "./chunk-PPYSEKFM.js";
40
+ import {
41
+ createMind2WebAdapter
42
+ } from "./chunk-TBKU5XQI.js";
43
+ import {
44
+ createNoMiraclAdapter
45
+ } from "./chunk-TSWPNOYM.js";
46
+ import {
47
+ createOpenRagBenchAdapter
48
+ } from "./chunk-KDIKRJGB.js";
49
+ import {
50
+ createProgrambenchAdapter
51
+ } from "./chunk-ZEWMTR5M.js";
52
+ import {
53
+ createCadBenchAdapter
54
+ } from "./chunk-PUIRNYI7.js";
55
+ import {
56
+ createCadGenBenchAdapter
57
+ } from "./chunk-C7T7WEK2.js";
58
+ import {
59
+ createCommit0Adapter
60
+ } from "./chunk-2XU6OGEN.js";
61
+ import {
62
+ createCragAdapter
63
+ } from "./chunk-HHXFIHXC.js";
64
+ import {
65
+ createDabstepAdapter
66
+ } from "./chunk-J3KDJNX2.js";
67
+ import {
68
+ createEnterpriseOpsGymAdapter
69
+ } from "./chunk-X5YKXC6V.js";
70
+ import {
71
+ createFinResearchBenchAdapter
72
+ } from "./chunk-YCGY7UIZ.js";
73
+ import {
74
+ createFinsearchcompAdapter
75
+ } from "./chunk-UPAMRDX4.js";
76
+ import {
77
+ createAecBenchAdapter
78
+ } from "./chunk-ODT47UAY.js";
79
+ import {
80
+ createAgentBenchAdapter
81
+ } from "./chunk-7WSD27QQ.js";
82
+ import {
83
+ createAppWorldAdapter,
84
+ createAppWorldReactAdapter
85
+ } from "./chunk-SYDW647C.js";
86
+ import {
87
+ createBfclAdapter
88
+ } from "./chunk-IFAV6KEM.js";
89
+ import {
90
+ createCadDesignAdapter
91
+ } from "./chunk-53UPUNBZ.js";
92
+
93
+ // src/adapters.ts
94
+ var ADAPTERS = {
95
+ "swe-bench": createSweBenchAdapter,
96
+ "terminal-bench": createTerminalBenchAdapter,
97
+ // Code-benches sharing ./benchmarks/_harness (stage → external evaluator → report).
98
+ // loadTasks fetches the REAL dataset (committed fixtures fallback offline); judge
99
+ // delegates to the benchmark's own harness and fails loud when it/Docker is absent.
100
+ "aec-bench": createAecBenchAdapter,
101
+ commit0: createCommit0Adapter,
102
+ dabstep: createDabstepAdapter,
103
+ programbench: createProgrambenchAdapter,
104
+ "webarena-verified": createWebArenaVerifiedAdapter,
105
+ "tau2-bench": createTau2BenchAdapter,
106
+ "tau3-banking": createTau3BankingAdapter,
107
+ agentbench: createAgentBenchAdapter,
108
+ bfcl: createBfclAdapter,
109
+ toollm: createToolLlmAdapter,
110
+ appworld: createAppWorldAdapter,
111
+ // AppWorld's native interactive protocol — the worker is the in-engine ReAct
112
+ // episode (execution feedback every turn), the mode published baselines use.
113
+ "appworld-react": createAppWorldReactAdapter,
114
+ "enterpriseops-gym": createEnterpriseOpsGymAdapter,
115
+ "cad-design": createCadDesignAdapter,
116
+ cadbench: createCadBenchAdapter,
117
+ cadgenbench: createCadGenBenchAdapter,
118
+ frames: createFramesAdapter,
119
+ ragbench: createRagBenchAdapter,
120
+ crag: createCragAdapter,
121
+ nomiracl: createNoMiraclAdapter,
122
+ "open-rag-bench": createOpenRagBenchAdapter,
123
+ "t2-ragbench": createT2RagBenchAdapter,
124
+ finresearchbench: createFinResearchBenchAdapter,
125
+ finsearchcomp: createFinsearchcompAdapter,
126
+ simpleqa: createSimpleQaAdapter,
127
+ hotpotqa: createHotpotqaAdapter,
128
+ // Deployable-checker code domain: worker = router completion, judge = Docker test
129
+ // run (--network=none). The steering A/B counterpart to humaneval-gate.mts (selection).
130
+ humaneval: createHumanEvalAdapter,
131
+ mind2web: createMind2WebAdapter,
132
+ "trata-hedge": createTrataHedgeAdapter
133
+ };
134
+ function resolveAdapter(key) {
135
+ const make = ADAPTERS[key];
136
+ if (!make) throw new Error(`unknown benchmark ${JSON.stringify(key)} (have: ${Object.keys(ADAPTERS).join(", ")})`);
137
+ return make();
138
+ }
139
+
140
+ export {
141
+ ADAPTERS,
142
+ resolveAdapter
143
+ };
144
+ //# sourceMappingURL=chunk-5SBJCB6W.js.map
@@ -0,0 +1 @@
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: ${Object.keys(ADAPTERS).join(', ')})`)\n return make()\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAuCO,IAAM,WAAmD;AAAA,EAC9D,aAAa;AAAA,EACb,kBAAkB;AAAA;AAAA;AAAA;AAAA,EAIlB,aAAa;AAAA,EACb,SAAS;AAAA,EACT,SAAS;AAAA,EACT,cAAc;AAAA,EACd,qBAAqB;AAAA,EACrB,cAAc;AAAA,EACd,gBAAgB;AAAA,EAChB,YAAY;AAAA,EACZ,MAAM;AAAA,EACN,QAAQ;AAAA,EACR,UAAU;AAAA;AAAA;AAAA,EAGV,kBAAkB;AAAA,EAClB,qBAAqB;AAAA,EACrB,cAAc;AAAA,EACd,UAAU;AAAA,EACV,aAAa;AAAA,EACb,QAAQ;AAAA,EACR,UAAU;AAAA,EACV,MAAM;AAAA,EACN,UAAU;AAAA,EACV,kBAAkB;AAAA,EAClB,eAAe;AAAA,EACf,kBAAkB;AAAA,EAClB,eAAe;AAAA,EACf,UAAU;AAAA,EACV,UAAU;AAAA;AAAA;AAAA,EAGV,WAAW;AAAA,EACX,UAAU;AAAA,EACV,eAAe;AACjB;AAGO,SAAS,eAAe,KAA+B;AAC5D,QAAM,OAAO,SAAS,GAAG;AACzB,MAAI,CAAC,KAAM,OAAM,IAAI,MAAM,qBAAqB,KAAK,UAAU,GAAG,CAAC,WAAW,OAAO,KAAK,QAAQ,EAAE,KAAK,IAAI,CAAC,GAAG;AACjH,SAAO,KAAK;AACd;","names":[]}
@@ -0,0 +1,118 @@
1
+ import {
2
+ benchRoot
3
+ } from "./chunk-LRRD7NAG.js";
4
+
5
+ // src/benchmarks/agentbench.ts
6
+ import { readFile } from "fs/promises";
7
+ import { join } from "path";
8
+ var FIXTURES = join(benchRoot, "fixtures", "agentbench-dbbench.json");
9
+ var DEFAULT_SPLIT = "dev";
10
+ var agentbenchDir = () => process.env.AGENTBENCH_DIR;
11
+ var agentbenchAnswerOutput = {
12
+ parse(events) {
13
+ let text = "";
14
+ for (const ev of events) {
15
+ const d = ev?.data;
16
+ const t = d?.finalText ?? d?.text ?? d?.result;
17
+ if (typeof t === "string" && t.length > 0) text = t;
18
+ }
19
+ const fences = [...text.matchAll(/```(?:text|answer)?\s*\n([\s\S]*?)```/g)];
20
+ return (fences.at(-1)?.[1] ?? text).trim();
21
+ }
22
+ };
23
+ function normalizeAnswer(value) {
24
+ return value.toLowerCase().replace(/\s+/g, " ").replace(/^["'`]+|["'`]+$/g, "").trim();
25
+ }
26
+ function rowToTask(row, index, split) {
27
+ const columns = row.table?.table_info?.columns?.map((c) => `${c.name}${c.type ? ` (${c.type})` : ""}`).join(", ");
28
+ const sampleRows = row.table?.table_info?.rows?.slice(0, 40);
29
+ const meta = {
30
+ labels: row.label,
31
+ split,
32
+ subset: "dbbench",
33
+ table: row.table
34
+ };
35
+ return {
36
+ id: `dbbench-${split}-${index}`,
37
+ split,
38
+ prompt: [
39
+ "Answer this AgentBench DBBench question using the table below.",
40
+ "Return only the answer value.",
41
+ "",
42
+ `Question: ${row.description}`,
43
+ row.table?.table_name ? `Table: ${row.table.table_name}` : void 0,
44
+ columns ? `Columns: ${columns}` : void 0,
45
+ sampleRows ? `Rows JSON: ${JSON.stringify(sampleRows)}` : void 0
46
+ ].filter(Boolean).join("\n"),
47
+ metadata: meta
48
+ };
49
+ }
50
+ function readMeta(task) {
51
+ const md = task.metadata;
52
+ if (!md || !Array.isArray(md.labels)) {
53
+ throw new Error(`agentbench task ${task.id} missing metadata \u2014 loadTasks did not populate it`);
54
+ }
55
+ return md;
56
+ }
57
+ function selectRows(rows, opts, split) {
58
+ let tasks = rows.map((row, index) => rowToTask(row, index, split));
59
+ if (opts.ids) {
60
+ const want = new Set(opts.ids);
61
+ tasks = tasks.filter((task) => want.has(task.id));
62
+ } else if (opts.limit !== void 0) {
63
+ tasks = tasks.slice(0, opts.limit);
64
+ }
65
+ if (tasks.length === 0) throw new Error(`AgentBench DBBench: no tasks matched ${JSON.stringify(opts)}`);
66
+ return tasks;
67
+ }
68
+ async function loadJsonl(path) {
69
+ const raw = await readFile(path, "utf8");
70
+ return raw.split("\n").map((line) => line.trim()).filter(Boolean).map((line) => JSON.parse(line));
71
+ }
72
+ async function loadFixtures(opts, split) {
73
+ const rows = JSON.parse(await readFile(FIXTURES, "utf8"));
74
+ console.warn(`[agentbench] AGENTBENCH_FIXTURES=1 \u2014 loading ${rows.length} DBBench adapter fixtures`);
75
+ return selectRows(rows, opts, split);
76
+ }
77
+ function createAgentBenchAdapter() {
78
+ const fixturesMode = process.env.AGENTBENCH_FIXTURES === "1";
79
+ return {
80
+ name: "agentbench",
81
+ output: agentbenchAnswerOutput,
82
+ async preflight() {
83
+ if (fixturesMode) return;
84
+ const dir = agentbenchDir();
85
+ if (!dir) {
86
+ throw new Error("AGENTBENCH_DIR is required. Fix: clone https://github.com/THUDM/AgentBench and set AGENTBENCH_DIR=/path/to/AgentBench.");
87
+ }
88
+ await loadJsonl(join(dir, "data", "dbbench", `${DEFAULT_SPLIT}.jsonl`));
89
+ },
90
+ async loadTasks(opts = {}) {
91
+ const split = opts.split ?? DEFAULT_SPLIT;
92
+ if (fixturesMode) return loadFixtures(opts, split);
93
+ const dir = agentbenchDir();
94
+ if (!dir) throw new Error("AGENTBENCH_DIR is required to load AgentBench DBBench tasks");
95
+ return selectRows(await loadJsonl(join(dir, "data", "dbbench", `${split}.jsonl`)), opts, split);
96
+ },
97
+ async goldArtifact(task) {
98
+ return readMeta(task).labels[0];
99
+ },
100
+ async judge(task, artifact) {
101
+ const meta = readMeta(task);
102
+ const got = normalizeAnswer(artifact);
103
+ const expected = meta.labels.map(normalizeAnswer);
104
+ const resolved = expected.includes(got);
105
+ return {
106
+ resolved,
107
+ score: resolved ? 1 : 0,
108
+ detail: JSON.stringify({ subset: meta.subset, split: meta.split, expected: meta.labels, got: artifact })
109
+ };
110
+ }
111
+ };
112
+ }
113
+
114
+ export {
115
+ agentbenchAnswerOutput,
116
+ createAgentBenchAdapter
117
+ };
118
+ //# sourceMappingURL=chunk-7WSD27QQ.js.map