@prismatic-io/lux 0.0.2-preview.4 → 0.0.2-preview.41

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 (756) hide show
  1. package/lib/answerers/claude-code/index.d.ts +15 -3
  2. package/lib/answerers/claude-code/index.d.ts.map +1 -1
  3. package/lib/answerers/claude-code/index.js +6 -5
  4. package/lib/answerers/claude-code/index.js.map +1 -1
  5. package/lib/answerers/persona/index.d.ts +44 -5
  6. package/lib/answerers/persona/index.d.ts.map +1 -1
  7. package/lib/answerers/persona/index.js +65 -34
  8. package/lib/answerers/persona/index.js.map +1 -1
  9. package/lib/answerers/terminal/index.d.ts +9 -3
  10. package/lib/answerers/terminal/index.d.ts.map +1 -1
  11. package/lib/answerers/terminal/index.js +6 -4
  12. package/lib/answerers/terminal/index.js.map +1 -1
  13. package/lib/assertions/authoring.d.ts +2 -2
  14. package/lib/assertions/authoring.d.ts.map +1 -1
  15. package/lib/assertions/authoring.js.map +1 -1
  16. package/lib/assertions/core/command-exits-zero.d.ts +5 -2
  17. package/lib/assertions/core/command-exits-zero.d.ts.map +1 -1
  18. package/lib/assertions/core/command-exits-zero.js +59 -14
  19. package/lib/assertions/core/command-exits-zero.js.map +1 -1
  20. package/lib/assertions/core/event-checks.d.ts +3 -0
  21. package/lib/assertions/core/event-checks.d.ts.map +1 -1
  22. package/lib/assertions/core/event-checks.js +49 -3
  23. package/lib/assertions/core/event-checks.js.map +1 -1
  24. package/lib/assertions/core/file-checks.d.ts +6 -0
  25. package/lib/assertions/core/file-checks.d.ts.map +1 -1
  26. package/lib/assertions/core/index.d.ts.map +1 -1
  27. package/lib/assertions/core/index.js +2 -1
  28. package/lib/assertions/core/index.js.map +1 -1
  29. package/lib/assertions/core/json-pointer-equals.d.ts +1 -0
  30. package/lib/assertions/core/json-pointer-equals.d.ts.map +1 -1
  31. package/lib/assertions/core/output-checks.d.ts +3 -0
  32. package/lib/assertions/core/output-checks.d.ts.map +1 -1
  33. package/lib/assertions/core/predicate.d.ts +1 -0
  34. package/lib/assertions/core/predicate.d.ts.map +1 -1
  35. package/lib/assertions/core/resource-checks.d.ts +25 -0
  36. package/lib/assertions/core/resource-checks.d.ts.map +1 -1
  37. package/lib/assertions/core/resource-checks.js +35 -5
  38. package/lib/assertions/core/resource-checks.js.map +1 -1
  39. package/lib/assertions/core/safe-regex.d.ts.map +1 -1
  40. package/lib/assertions/core/safe-regex.js +7 -1
  41. package/lib/assertions/core/safe-regex.js.map +1 -1
  42. package/lib/assertions/rubric/evidence.d.ts +88 -0
  43. package/lib/assertions/rubric/evidence.d.ts.map +1 -0
  44. package/lib/assertions/rubric/evidence.js +75 -0
  45. package/lib/assertions/rubric/evidence.js.map +1 -0
  46. package/lib/assertions/rubric/index.d.ts +88 -34
  47. package/lib/assertions/rubric/index.d.ts.map +1 -1
  48. package/lib/assertions/rubric/index.js +67 -43
  49. package/lib/assertions/rubric/index.js.map +1 -1
  50. package/lib/assertions/rubric/internal.d.ts +30 -5
  51. package/lib/assertions/rubric/internal.d.ts.map +1 -1
  52. package/lib/assertions/rubric/internal.js +211 -46
  53. package/lib/assertions/rubric/internal.js.map +1 -1
  54. package/lib/authoring.d.ts +37 -2
  55. package/lib/authoring.d.ts.map +1 -1
  56. package/lib/authoring.js +36 -1
  57. package/lib/authoring.js.map +1 -1
  58. package/lib/cli/bin.js +2 -13
  59. package/lib/cli/bin.js.map +1 -1
  60. package/lib/cli/command-runtime.d.ts +3 -3
  61. package/lib/cli/command-runtime.d.ts.map +1 -1
  62. package/lib/cli/command-runtime.js +17 -6
  63. package/lib/cli/command-runtime.js.map +1 -1
  64. package/lib/cli/doctor.d.ts +55 -0
  65. package/lib/cli/doctor.d.ts.map +1 -0
  66. package/lib/cli/doctor.js +14 -0
  67. package/lib/cli/doctor.js.map +1 -0
  68. package/lib/cli/harness-plan.d.ts +27 -0
  69. package/lib/cli/harness-plan.d.ts.map +1 -0
  70. package/lib/cli/harness-plan.js +75 -0
  71. package/lib/cli/harness-plan.js.map +1 -0
  72. package/lib/cli/init-templates.d.ts +11 -0
  73. package/lib/cli/init-templates.d.ts.map +1 -0
  74. package/lib/cli/init-templates.js +153 -0
  75. package/lib/cli/init-templates.js.map +1 -0
  76. package/lib/cli/init.d.ts +10 -0
  77. package/lib/cli/init.d.ts.map +1 -1
  78. package/lib/cli/init.js +34 -173
  79. package/lib/cli/init.js.map +1 -1
  80. package/lib/cli/output-schemas.d.ts +1412 -0
  81. package/lib/cli/output-schemas.d.ts.map +1 -0
  82. package/lib/cli/output-schemas.js +94 -0
  83. package/lib/cli/output-schemas.js.map +1 -0
  84. package/lib/cli/program.d.ts +140 -12
  85. package/lib/cli/program.d.ts.map +1 -1
  86. package/lib/cli/program.js +783 -650
  87. package/lib/cli/program.js.map +1 -1
  88. package/lib/cli/render/format.d.ts.map +1 -1
  89. package/lib/cli/render/format.js +1 -1
  90. package/lib/cli/render/format.js.map +1 -1
  91. package/lib/cli/render/reporter.d.ts.map +1 -1
  92. package/lib/cli/render/reporter.js +35 -8
  93. package/lib/cli/render/reporter.js.map +1 -1
  94. package/lib/cli/run-options.d.ts +19 -4
  95. package/lib/cli/run-options.d.ts.map +1 -1
  96. package/lib/cli/run-options.js +100 -19
  97. package/lib/cli/run-options.js.map +1 -1
  98. package/lib/cli/view.d.ts +15 -0
  99. package/lib/cli/view.d.ts.map +1 -0
  100. package/lib/cli/view.js +168 -0
  101. package/lib/cli/view.js.map +1 -0
  102. package/lib/container/driver.d.ts +27 -0
  103. package/lib/container/driver.d.ts.map +1 -0
  104. package/lib/container/driver.js +199 -0
  105. package/lib/container/driver.js.map +1 -0
  106. package/lib/container/paths.d.ts +3 -0
  107. package/lib/container/paths.d.ts.map +1 -0
  108. package/lib/container/paths.js +27 -0
  109. package/lib/container/paths.js.map +1 -0
  110. package/lib/container/protocol.d.ts +12 -0
  111. package/lib/container/protocol.d.ts.map +1 -0
  112. package/lib/container/protocol.js +38 -0
  113. package/lib/container/protocol.js.map +1 -0
  114. package/lib/container/resources.d.ts +14 -0
  115. package/lib/container/resources.d.ts.map +1 -0
  116. package/lib/container/resources.js +141 -0
  117. package/lib/container/resources.js.map +1 -0
  118. package/lib/container/worker.d.ts +2 -0
  119. package/lib/container/worker.d.ts.map +1 -0
  120. package/lib/container/worker.js +85 -0
  121. package/lib/container/worker.js.map +1 -0
  122. package/lib/core/action-events.d.ts +29 -0
  123. package/lib/core/action-events.d.ts.map +1 -0
  124. package/lib/core/action-events.js +51 -0
  125. package/lib/core/action-events.js.map +1 -0
  126. package/lib/core/agent-activity.d.ts +49 -0
  127. package/lib/core/agent-activity.d.ts.map +1 -0
  128. package/lib/core/agent-activity.js +206 -0
  129. package/lib/core/agent-activity.js.map +1 -0
  130. package/lib/core/agent-environment.d.ts +5 -0
  131. package/lib/core/agent-environment.d.ts.map +1 -0
  132. package/lib/core/agent-environment.js +47 -0
  133. package/lib/core/agent-environment.js.map +1 -0
  134. package/lib/core/agent-events.d.ts +26 -0
  135. package/lib/core/agent-events.d.ts.map +1 -0
  136. package/lib/core/agent-events.js +13 -0
  137. package/lib/core/agent-events.js.map +1 -0
  138. package/lib/core/annotation.d.ts +6 -2
  139. package/lib/core/annotation.d.ts.map +1 -1
  140. package/lib/core/annotation.js +2 -1
  141. package/lib/core/annotation.js.map +1 -1
  142. package/lib/core/answerer.d.ts +12 -0
  143. package/lib/core/answerer.d.ts.map +1 -1
  144. package/lib/core/answerer.js +2 -0
  145. package/lib/core/answerer.js.map +1 -1
  146. package/lib/core/artifact-evidence.d.ts +6 -1
  147. package/lib/core/artifact-evidence.d.ts.map +1 -1
  148. package/lib/core/artifact-evidence.js +27 -6
  149. package/lib/core/artifact-evidence.js.map +1 -1
  150. package/lib/core/assertion.d.ts +10 -4
  151. package/lib/core/assertion.d.ts.map +1 -1
  152. package/lib/core/assertion.js +8 -2
  153. package/lib/core/assertion.js.map +1 -1
  154. package/lib/core/case-defaults.d.ts +16 -0
  155. package/lib/core/case-defaults.d.ts.map +1 -0
  156. package/lib/core/case-defaults.js +42 -0
  157. package/lib/core/case-defaults.js.map +1 -0
  158. package/lib/core/case.d.ts +120 -6
  159. package/lib/core/case.d.ts.map +1 -1
  160. package/lib/core/case.js +51 -5
  161. package/lib/core/case.js.map +1 -1
  162. package/lib/core/command-verification.d.ts +10 -0
  163. package/lib/core/command-verification.d.ts.map +1 -0
  164. package/lib/core/command-verification.js +66 -0
  165. package/lib/core/command-verification.js.map +1 -0
  166. package/lib/core/driver-execution.d.ts +83 -0
  167. package/lib/core/driver-execution.d.ts.map +1 -0
  168. package/lib/core/driver-execution.js +72 -0
  169. package/lib/core/driver-execution.js.map +1 -0
  170. package/lib/core/driver-metrics.d.ts +98 -0
  171. package/lib/core/driver-metrics.d.ts.map +1 -0
  172. package/lib/core/driver-metrics.js +36 -0
  173. package/lib/core/driver-metrics.js.map +1 -0
  174. package/lib/core/driver.d.ts +64 -0
  175. package/lib/core/driver.d.ts.map +1 -1
  176. package/lib/core/driver.js +28 -0
  177. package/lib/core/driver.js.map +1 -1
  178. package/lib/core/duration.d.ts +5 -0
  179. package/lib/core/duration.d.ts.map +1 -0
  180. package/lib/core/duration.js +49 -0
  181. package/lib/core/duration.js.map +1 -0
  182. package/lib/core/environment.d.ts +18 -0
  183. package/lib/core/environment.d.ts.map +1 -0
  184. package/lib/core/environment.js +21 -0
  185. package/lib/core/environment.js.map +1 -0
  186. package/lib/core/evaluation-clusters.d.ts +17 -0
  187. package/lib/core/evaluation-clusters.d.ts.map +1 -0
  188. package/lib/core/evaluation-clusters.js +23 -0
  189. package/lib/core/evaluation-clusters.js.map +1 -0
  190. package/lib/core/events.d.ts +34 -0
  191. package/lib/core/events.d.ts.map +1 -1
  192. package/lib/core/events.js +11 -2
  193. package/lib/core/events.js.map +1 -1
  194. package/lib/core/experiment.d.ts +33 -2
  195. package/lib/core/experiment.d.ts.map +1 -1
  196. package/lib/core/experiment.js +40 -19
  197. package/lib/core/experiment.js.map +1 -1
  198. package/lib/core/file-tree.d.ts +7 -0
  199. package/lib/core/file-tree.d.ts.map +1 -0
  200. package/lib/core/file-tree.js +35 -0
  201. package/lib/core/file-tree.js.map +1 -0
  202. package/lib/core/global-id.d.ts +6 -0
  203. package/lib/core/global-id.d.ts.map +1 -0
  204. package/lib/core/global-id.js +17 -0
  205. package/lib/core/global-id.js.map +1 -0
  206. package/lib/core/harness-catalog.d.ts +35 -0
  207. package/lib/core/harness-catalog.d.ts.map +1 -0
  208. package/lib/core/harness-catalog.js +45 -0
  209. package/lib/core/harness-catalog.js.map +1 -0
  210. package/lib/core/harness-catalogs/antigravity.d.ts +11 -0
  211. package/lib/core/harness-catalogs/antigravity.d.ts.map +1 -0
  212. package/lib/core/harness-catalogs/antigravity.js +23 -0
  213. package/lib/core/harness-catalogs/antigravity.js.map +1 -0
  214. package/lib/core/harness-catalogs/claude-code.d.ts +12 -0
  215. package/lib/core/harness-catalogs/claude-code.d.ts.map +1 -0
  216. package/lib/core/harness-catalogs/claude-code.js +24 -0
  217. package/lib/core/harness-catalogs/claude-code.js.map +1 -0
  218. package/lib/core/harness-catalogs/codex.d.ts +11 -0
  219. package/lib/core/harness-catalogs/codex.d.ts.map +1 -0
  220. package/lib/core/harness-catalogs/codex.js +26 -0
  221. package/lib/core/harness-catalogs/codex.js.map +1 -0
  222. package/lib/core/harness-catalogs/copilot.d.ts +29 -0
  223. package/lib/core/harness-catalogs/copilot.d.ts.map +1 -0
  224. package/lib/core/harness-catalogs/copilot.js +45 -0
  225. package/lib/core/harness-catalogs/copilot.js.map +1 -0
  226. package/lib/core/harness-catalogs/cursor.d.ts +2 -0
  227. package/lib/core/harness-catalogs/cursor.d.ts.map +1 -0
  228. package/lib/core/harness-catalogs/cursor.js +13 -0
  229. package/lib/core/harness-catalogs/cursor.js.map +1 -0
  230. package/lib/core/harness-catalogs/grok-build.d.ts +7 -0
  231. package/lib/core/harness-catalogs/grok-build.d.ts.map +1 -0
  232. package/lib/core/harness-catalogs/grok-build.js +18 -0
  233. package/lib/core/harness-catalogs/grok-build.js.map +1 -0
  234. package/lib/core/harness-catalogs/index.d.ts +73 -0
  235. package/lib/core/harness-catalogs/index.d.ts.map +1 -0
  236. package/lib/core/harness-catalogs/index.js +17 -0
  237. package/lib/core/harness-catalogs/index.js.map +1 -0
  238. package/lib/core/harness-matrix.d.ts +154 -0
  239. package/lib/core/harness-matrix.d.ts.map +1 -0
  240. package/lib/core/harness-matrix.js +204 -0
  241. package/lib/core/harness-matrix.js.map +1 -0
  242. package/lib/core/index.d.ts +12 -0
  243. package/lib/core/index.d.ts.map +1 -1
  244. package/lib/core/index.js +12 -0
  245. package/lib/core/index.js.map +1 -1
  246. package/lib/core/lifecycle-fixtures.d.ts.map +1 -1
  247. package/lib/core/lifecycle-fixtures.js +4 -2
  248. package/lib/core/lifecycle-fixtures.js.map +1 -1
  249. package/lib/core/model-cli.d.ts +31 -15
  250. package/lib/core/model-cli.d.ts.map +1 -1
  251. package/lib/core/model-cli.js +53 -23
  252. package/lib/core/model-cli.js.map +1 -1
  253. package/lib/core/platform-process.d.ts +7 -2
  254. package/lib/core/platform-process.d.ts.map +1 -1
  255. package/lib/core/platform-process.js +54 -16
  256. package/lib/core/platform-process.js.map +1 -1
  257. package/lib/core/reference-defaults.d.ts +5 -0
  258. package/lib/core/reference-defaults.d.ts.map +1 -0
  259. package/lib/core/reference-defaults.js +42 -0
  260. package/lib/core/reference-defaults.js.map +1 -0
  261. package/lib/core/rubric.d.ts +1 -0
  262. package/lib/core/rubric.d.ts.map +1 -1
  263. package/lib/core/run.d.ts +292 -1
  264. package/lib/core/run.d.ts.map +1 -1
  265. package/lib/core/run.js +27 -0
  266. package/lib/core/run.js.map +1 -1
  267. package/lib/core/secure-copy.d.ts.map +1 -0
  268. package/lib/core/secure-copy.js.map +1 -0
  269. package/lib/core/usage.d.ts +5 -0
  270. package/lib/core/usage.d.ts.map +1 -1
  271. package/lib/core/usage.js +3 -0
  272. package/lib/core/usage.js.map +1 -1
  273. package/lib/core/wire.d.ts +11 -0
  274. package/lib/core/wire.d.ts.map +1 -1
  275. package/lib/drivers/antigravity/artifacts.d.ts +2 -0
  276. package/lib/drivers/antigravity/artifacts.d.ts.map +1 -0
  277. package/lib/drivers/antigravity/artifacts.js +33 -0
  278. package/lib/drivers/antigravity/artifacts.js.map +1 -0
  279. package/lib/drivers/antigravity/config.d.ts +88 -0
  280. package/lib/drivers/antigravity/config.d.ts.map +1 -0
  281. package/lib/drivers/antigravity/config.js +87 -0
  282. package/lib/drivers/antigravity/config.js.map +1 -0
  283. package/lib/drivers/antigravity/events.d.ts +4 -0
  284. package/lib/drivers/antigravity/events.d.ts.map +1 -0
  285. package/lib/drivers/antigravity/events.js +241 -0
  286. package/lib/drivers/antigravity/events.js.map +1 -0
  287. package/lib/drivers/antigravity/index.d.ts +114 -0
  288. package/lib/drivers/antigravity/index.d.ts.map +1 -0
  289. package/lib/drivers/antigravity/index.js +113 -0
  290. package/lib/drivers/antigravity/index.js.map +1 -0
  291. package/lib/drivers/antigravity/runtime.d.ts +7 -0
  292. package/lib/drivers/antigravity/runtime.d.ts.map +1 -0
  293. package/lib/drivers/antigravity/runtime.js +83 -0
  294. package/lib/drivers/antigravity/runtime.js.map +1 -0
  295. package/lib/drivers/claude-code/deferred-process.d.ts +17 -0
  296. package/lib/drivers/claude-code/deferred-process.d.ts.map +1 -0
  297. package/lib/drivers/claude-code/deferred-process.js +108 -0
  298. package/lib/drivers/claude-code/deferred-process.js.map +1 -0
  299. package/lib/drivers/claude-code/index.d.ts +187 -10
  300. package/lib/drivers/claude-code/index.d.ts.map +1 -1
  301. package/lib/drivers/claude-code/index.js +198 -253
  302. package/lib/drivers/claude-code/index.js.map +1 -1
  303. package/lib/drivers/claude-code/isolation.d.ts +8 -0
  304. package/lib/drivers/claude-code/isolation.d.ts.map +1 -0
  305. package/lib/drivers/claude-code/isolation.js +199 -0
  306. package/lib/drivers/claude-code/isolation.js.map +1 -0
  307. package/lib/drivers/claude-code/parse-events.d.ts.map +1 -1
  308. package/lib/drivers/claude-code/parse-events.js +73 -8
  309. package/lib/drivers/claude-code/parse-events.js.map +1 -1
  310. package/lib/drivers/codex/app-events.d.ts +9 -2
  311. package/lib/drivers/codex/app-events.d.ts.map +1 -1
  312. package/lib/drivers/codex/app-events.js +104 -12
  313. package/lib/drivers/codex/app-events.js.map +1 -1
  314. package/lib/drivers/codex/config.d.ts +69 -5
  315. package/lib/drivers/codex/config.d.ts.map +1 -1
  316. package/lib/drivers/codex/config.js +48 -8
  317. package/lib/drivers/codex/config.js.map +1 -1
  318. package/lib/drivers/codex/exec-events.d.ts +4 -0
  319. package/lib/drivers/codex/exec-events.d.ts.map +1 -0
  320. package/lib/drivers/codex/exec-events.js +230 -0
  321. package/lib/drivers/codex/exec-events.js.map +1 -0
  322. package/lib/drivers/codex/index.d.ts +142 -4
  323. package/lib/drivers/codex/index.d.ts.map +1 -1
  324. package/lib/drivers/codex/index.js +114 -276
  325. package/lib/drivers/codex/index.js.map +1 -1
  326. package/lib/drivers/copilot/config.d.ts +84 -0
  327. package/lib/drivers/copilot/config.d.ts.map +1 -0
  328. package/lib/drivers/copilot/config.js +57 -0
  329. package/lib/drivers/copilot/config.js.map +1 -0
  330. package/lib/drivers/copilot/index.d.ts +107 -0
  331. package/lib/drivers/copilot/index.d.ts.map +1 -0
  332. package/lib/drivers/copilot/index.js +268 -0
  333. package/lib/drivers/copilot/index.js.map +1 -0
  334. package/lib/drivers/cursor/config.d.ts +28 -0
  335. package/lib/drivers/cursor/config.d.ts.map +1 -0
  336. package/lib/drivers/cursor/config.js +28 -0
  337. package/lib/drivers/cursor/config.js.map +1 -0
  338. package/lib/drivers/cursor/index.d.ts +44 -0
  339. package/lib/drivers/cursor/index.d.ts.map +1 -0
  340. package/lib/drivers/cursor/index.js +286 -0
  341. package/lib/drivers/cursor/index.js.map +1 -0
  342. package/lib/drivers/cursor/interaction.d.ts +6 -0
  343. package/lib/drivers/cursor/interaction.d.ts.map +1 -0
  344. package/lib/drivers/cursor/interaction.js +55 -0
  345. package/lib/drivers/cursor/interaction.js.map +1 -0
  346. package/lib/drivers/grok-build/config.d.ts +68 -0
  347. package/lib/drivers/grok-build/config.d.ts.map +1 -0
  348. package/lib/drivers/grok-build/config.js +40 -0
  349. package/lib/drivers/grok-build/config.js.map +1 -0
  350. package/lib/drivers/grok-build/index.d.ts +87 -0
  351. package/lib/drivers/grok-build/index.d.ts.map +1 -0
  352. package/lib/drivers/grok-build/index.js +293 -0
  353. package/lib/drivers/grok-build/index.js.map +1 -0
  354. package/lib/drivers/mcp/index.d.ts +37 -4
  355. package/lib/drivers/mcp/index.d.ts.map +1 -1
  356. package/lib/drivers/mcp/index.js +25 -11
  357. package/lib/drivers/mcp/index.js.map +1 -1
  358. package/lib/drivers/shared/acp-events.d.ts +15 -0
  359. package/lib/drivers/shared/acp-events.d.ts.map +1 -0
  360. package/lib/drivers/shared/acp-events.js +83 -0
  361. package/lib/drivers/shared/acp-events.js.map +1 -0
  362. package/lib/drivers/shared/acp-interaction.d.ts +7 -0
  363. package/lib/drivers/shared/acp-interaction.d.ts.map +1 -0
  364. package/lib/drivers/shared/acp-interaction.js +20 -0
  365. package/lib/drivers/shared/acp-interaction.js.map +1 -0
  366. package/lib/drivers/shared/acp-session.d.ts +4 -0
  367. package/lib/drivers/shared/acp-session.d.ts.map +1 -0
  368. package/lib/drivers/shared/acp-session.js +23 -0
  369. package/lib/drivers/shared/acp-session.js.map +1 -0
  370. package/lib/drivers/shared/acp-transport.d.ts +40 -0
  371. package/lib/drivers/shared/acp-transport.d.ts.map +1 -0
  372. package/lib/drivers/shared/acp-transport.js +152 -0
  373. package/lib/drivers/shared/acp-transport.js.map +1 -0
  374. package/lib/drivers/shared/artifacts.d.ts.map +1 -1
  375. package/lib/drivers/shared/artifacts.js +3 -2
  376. package/lib/drivers/shared/artifacts.js.map +1 -1
  377. package/lib/drivers/shared/experiment-behavior.d.ts +13 -0
  378. package/lib/drivers/shared/experiment-behavior.d.ts.map +1 -0
  379. package/lib/drivers/shared/experiment-behavior.js +200 -0
  380. package/lib/drivers/shared/experiment-behavior.js.map +1 -0
  381. package/lib/drivers/shared/experiment-policy.d.ts +4 -0
  382. package/lib/drivers/shared/experiment-policy.d.ts.map +1 -0
  383. package/lib/drivers/shared/experiment-policy.js +149 -0
  384. package/lib/drivers/shared/experiment-policy.js.map +1 -0
  385. package/lib/drivers/shared/skills.d.ts +3 -0
  386. package/lib/drivers/shared/skills.d.ts.map +1 -0
  387. package/lib/drivers/shared/skills.js +73 -0
  388. package/lib/drivers/shared/skills.js.map +1 -0
  389. package/lib/drivers/subprocess/index.d.ts +39 -4
  390. package/lib/drivers/subprocess/index.d.ts.map +1 -1
  391. package/lib/drivers/subprocess/index.js +56 -12
  392. package/lib/drivers/subprocess/index.js.map +1 -1
  393. package/lib/index.d.ts +24 -8
  394. package/lib/index.d.ts.map +1 -1
  395. package/lib/index.js +20 -5
  396. package/lib/index.js.map +1 -1
  397. package/lib/optimization/reflective.d.ts +3 -1
  398. package/lib/optimization/reflective.d.ts.map +1 -1
  399. package/lib/optimization/reflective.js +3 -2
  400. package/lib/optimization/reflective.js.map +1 -1
  401. package/lib/orchestrator/annotation-loader.d.ts +6 -2
  402. package/lib/orchestrator/annotation-loader.d.ts.map +1 -1
  403. package/lib/orchestrator/annotation-loader.js +8 -0
  404. package/lib/orchestrator/annotation-loader.js.map +1 -1
  405. package/lib/orchestrator/annotation-store.d.ts.map +1 -1
  406. package/lib/orchestrator/annotation-store.js +16 -2
  407. package/lib/orchestrator/annotation-store.js.map +1 -1
  408. package/lib/orchestrator/artifact-origin.d.ts +3 -0
  409. package/lib/orchestrator/artifact-origin.d.ts.map +1 -0
  410. package/lib/orchestrator/artifact-origin.js +40 -0
  411. package/lib/orchestrator/artifact-origin.js.map +1 -0
  412. package/lib/orchestrator/campaign-lifecycle.d.ts +1 -0
  413. package/lib/orchestrator/campaign-lifecycle.d.ts.map +1 -1
  414. package/lib/orchestrator/campaign-lifecycle.js +4 -1
  415. package/lib/orchestrator/campaign-lifecycle.js.map +1 -1
  416. package/lib/orchestrator/candidates.d.ts.map +1 -1
  417. package/lib/orchestrator/candidates.js +3 -2
  418. package/lib/orchestrator/candidates.js.map +1 -1
  419. package/lib/orchestrator/compare.d.ts +23 -6
  420. package/lib/orchestrator/compare.d.ts.map +1 -1
  421. package/lib/orchestrator/compare.js +45 -12
  422. package/lib/orchestrator/compare.js.map +1 -1
  423. package/lib/orchestrator/comparison-identity.d.ts.map +1 -1
  424. package/lib/orchestrator/comparison-identity.js +2 -0
  425. package/lib/orchestrator/comparison-identity.js.map +1 -1
  426. package/lib/orchestrator/config.d.ts +231 -10
  427. package/lib/orchestrator/config.d.ts.map +1 -1
  428. package/lib/orchestrator/config.js +21 -2
  429. package/lib/orchestrator/config.js.map +1 -1
  430. package/lib/orchestrator/discover.d.ts +1 -0
  431. package/lib/orchestrator/discover.d.ts.map +1 -1
  432. package/lib/orchestrator/discover.js +7 -1
  433. package/lib/orchestrator/discover.js.map +1 -1
  434. package/lib/orchestrator/doctor.d.ts +1 -0
  435. package/lib/orchestrator/doctor.d.ts.map +1 -1
  436. package/lib/orchestrator/doctor.js +163 -78
  437. package/lib/orchestrator/doctor.js.map +1 -1
  438. package/lib/orchestrator/driver-capabilities.d.ts +4 -0
  439. package/lib/orchestrator/driver-capabilities.d.ts.map +1 -0
  440. package/lib/orchestrator/driver-capabilities.js +13 -0
  441. package/lib/orchestrator/driver-capabilities.js.map +1 -0
  442. package/lib/orchestrator/driver-identity.d.ts +5 -0
  443. package/lib/orchestrator/driver-identity.d.ts.map +1 -0
  444. package/lib/orchestrator/driver-identity.js +51 -0
  445. package/lib/orchestrator/driver-identity.js.map +1 -0
  446. package/lib/orchestrator/experiment-context.d.ts +52 -4
  447. package/lib/orchestrator/experiment-context.d.ts.map +1 -1
  448. package/lib/orchestrator/experiment-context.js +2 -1
  449. package/lib/orchestrator/experiment-context.js.map +1 -1
  450. package/lib/orchestrator/experiment-contracts.d.ts +31 -1
  451. package/lib/orchestrator/experiment-contracts.d.ts.map +1 -1
  452. package/lib/orchestrator/experiment-corpus.d.ts +2 -1
  453. package/lib/orchestrator/experiment-corpus.d.ts.map +1 -1
  454. package/lib/orchestrator/experiment-corpus.js +32 -24
  455. package/lib/orchestrator/experiment-corpus.js.map +1 -1
  456. package/lib/orchestrator/experiment-effects.d.ts +33 -0
  457. package/lib/orchestrator/experiment-effects.d.ts.map +1 -0
  458. package/lib/orchestrator/experiment-effects.js +46 -0
  459. package/lib/orchestrator/experiment-effects.js.map +1 -0
  460. package/lib/orchestrator/experiment-evaluation.d.ts.map +1 -1
  461. package/lib/orchestrator/experiment-evaluation.js +70 -5
  462. package/lib/orchestrator/experiment-evaluation.js.map +1 -1
  463. package/lib/orchestrator/experiment-identity.d.ts +3 -0
  464. package/lib/orchestrator/experiment-identity.d.ts.map +1 -1
  465. package/lib/orchestrator/experiment-identity.js +50 -42
  466. package/lib/orchestrator/experiment-identity.js.map +1 -1
  467. package/lib/orchestrator/experiment-loader.d.ts +8 -3
  468. package/lib/orchestrator/experiment-loader.d.ts.map +1 -1
  469. package/lib/orchestrator/experiment-preflight.d.ts +3 -0
  470. package/lib/orchestrator/experiment-preflight.d.ts.map +1 -0
  471. package/lib/orchestrator/experiment-preflight.js +21 -0
  472. package/lib/orchestrator/experiment-preflight.js.map +1 -0
  473. package/lib/orchestrator/experiment-report.d.ts +91 -1
  474. package/lib/orchestrator/experiment-report.d.ts.map +1 -1
  475. package/lib/orchestrator/experiment-report.js +9 -1
  476. package/lib/orchestrator/experiment-report.js.map +1 -1
  477. package/lib/orchestrator/experiment-runs.d.ts +63 -2
  478. package/lib/orchestrator/experiment-runs.d.ts.map +1 -1
  479. package/lib/orchestrator/experiment-runs.js +13 -5
  480. package/lib/orchestrator/experiment-runs.js.map +1 -1
  481. package/lib/orchestrator/experiment-runtime-identity.d.ts +0 -16
  482. package/lib/orchestrator/experiment-runtime-identity.d.ts.map +1 -1
  483. package/lib/orchestrator/experiment-runtime-identity.js +5 -220
  484. package/lib/orchestrator/experiment-runtime-identity.js.map +1 -1
  485. package/lib/orchestrator/experiment-selection.d.ts.map +1 -1
  486. package/lib/orchestrator/experiment-selection.js +151 -43
  487. package/lib/orchestrator/experiment-selection.js.map +1 -1
  488. package/lib/orchestrator/experiment.d.ts +3 -3
  489. package/lib/orchestrator/experiment.d.ts.map +1 -1
  490. package/lib/orchestrator/experiment.js +26 -21
  491. package/lib/orchestrator/experiment.js.map +1 -1
  492. package/lib/orchestrator/fixture-identity.d.ts +1 -0
  493. package/lib/orchestrator/fixture-identity.d.ts.map +1 -1
  494. package/lib/orchestrator/fixture-identity.js +6 -1
  495. package/lib/orchestrator/fixture-identity.js.map +1 -1
  496. package/lib/orchestrator/fs-read.d.ts +1 -6
  497. package/lib/orchestrator/fs-read.d.ts.map +1 -1
  498. package/lib/orchestrator/fs-read.js +2 -32
  499. package/lib/orchestrator/fs-read.js.map +1 -1
  500. package/lib/orchestrator/grader-experiment.d.ts +23 -5
  501. package/lib/orchestrator/grader-experiment.d.ts.map +1 -1
  502. package/lib/orchestrator/grader-experiment.js +54 -9
  503. package/lib/orchestrator/grader-experiment.js.map +1 -1
  504. package/lib/orchestrator/grading.d.ts +1 -0
  505. package/lib/orchestrator/grading.d.ts.map +1 -1
  506. package/lib/orchestrator/grading.js +41 -33
  507. package/lib/orchestrator/grading.js.map +1 -1
  508. package/lib/orchestrator/harness-catalog-drift.d.ts +22 -0
  509. package/lib/orchestrator/harness-catalog-drift.d.ts.map +1 -0
  510. package/lib/orchestrator/harness-catalog-drift.js +90 -0
  511. package/lib/orchestrator/harness-catalog-drift.js.map +1 -0
  512. package/lib/orchestrator/harness-registry.d.ts +4 -0
  513. package/lib/orchestrator/harness-registry.d.ts.map +1 -0
  514. package/lib/orchestrator/harness-registry.js +5 -0
  515. package/lib/orchestrator/harness-registry.js.map +1 -0
  516. package/lib/orchestrator/index.d.ts +1 -1
  517. package/lib/orchestrator/index.d.ts.map +1 -1
  518. package/lib/orchestrator/index.js +1 -1
  519. package/lib/orchestrator/index.js.map +1 -1
  520. package/lib/orchestrator/list-runs.d.ts +8 -1
  521. package/lib/orchestrator/list-runs.d.ts.map +1 -1
  522. package/lib/orchestrator/list-runs.js +83 -18
  523. package/lib/orchestrator/list-runs.js.map +1 -1
  524. package/lib/orchestrator/optimization-lifecycle.d.ts.map +1 -1
  525. package/lib/orchestrator/optimization-lifecycle.js +3 -1
  526. package/lib/orchestrator/optimization-lifecycle.js.map +1 -1
  527. package/lib/orchestrator/orchestrator.d.ts +17 -2
  528. package/lib/orchestrator/orchestrator.d.ts.map +1 -1
  529. package/lib/orchestrator/orchestrator.js +59 -179
  530. package/lib/orchestrator/orchestrator.js.map +1 -1
  531. package/lib/orchestrator/paired-effects.d.ts +34 -0
  532. package/lib/orchestrator/paired-effects.d.ts.map +1 -0
  533. package/lib/orchestrator/paired-effects.js +164 -0
  534. package/lib/orchestrator/paired-effects.js.map +1 -0
  535. package/lib/orchestrator/profile.d.ts +4 -0
  536. package/lib/orchestrator/profile.d.ts.map +1 -0
  537. package/lib/orchestrator/profile.js +10 -0
  538. package/lib/orchestrator/profile.js.map +1 -0
  539. package/lib/orchestrator/provenance.d.ts +5 -1
  540. package/lib/orchestrator/provenance.d.ts.map +1 -1
  541. package/lib/orchestrator/provenance.js +16 -3
  542. package/lib/orchestrator/provenance.js.map +1 -1
  543. package/lib/orchestrator/regrade.d.ts +4 -2
  544. package/lib/orchestrator/regrade.d.ts.map +1 -1
  545. package/lib/orchestrator/regrade.js +5 -1
  546. package/lib/orchestrator/regrade.js.map +1 -1
  547. package/lib/orchestrator/rubric-loader.js +1 -1
  548. package/lib/orchestrator/rubric-loader.js.map +1 -1
  549. package/lib/orchestrator/run-dir.d.ts +2 -0
  550. package/lib/orchestrator/run-dir.d.ts.map +1 -1
  551. package/lib/orchestrator/run-dir.js +1 -1
  552. package/lib/orchestrator/run-dir.js.map +1 -1
  553. package/lib/orchestrator/run-execution.d.ts +3 -0
  554. package/lib/orchestrator/run-execution.d.ts.map +1 -1
  555. package/lib/orchestrator/run-execution.js +64 -10
  556. package/lib/orchestrator/run-execution.js.map +1 -1
  557. package/lib/orchestrator/run-lifecycle.d.ts +1 -0
  558. package/lib/orchestrator/run-lifecycle.d.ts.map +1 -1
  559. package/lib/orchestrator/run-lifecycle.js +3 -2
  560. package/lib/orchestrator/run-lifecycle.js.map +1 -1
  561. package/lib/orchestrator/run-workspace.d.ts +10 -0
  562. package/lib/orchestrator/run-workspace.d.ts.map +1 -0
  563. package/lib/orchestrator/run-workspace.js +58 -0
  564. package/lib/orchestrator/run-workspace.js.map +1 -0
  565. package/lib/orchestrator/suite-compare.d.ts +37 -0
  566. package/lib/orchestrator/suite-compare.d.ts.map +1 -1
  567. package/lib/orchestrator/suite-compare.js +44 -22
  568. package/lib/orchestrator/suite-compare.js.map +1 -1
  569. package/lib/orchestrator/suite-runs.d.ts +6 -1
  570. package/lib/orchestrator/suite-runs.d.ts.map +1 -1
  571. package/lib/orchestrator/suite-runs.js +14 -2
  572. package/lib/orchestrator/suite-runs.js.map +1 -1
  573. package/lib/orchestrator/suite-summary.d.ts +79 -0
  574. package/lib/orchestrator/suite-summary.d.ts.map +1 -1
  575. package/lib/orchestrator/suite-summary.js +67 -14
  576. package/lib/orchestrator/suite-summary.js.map +1 -1
  577. package/lib/previewer/artifacts.d.ts +25 -0
  578. package/lib/previewer/artifacts.d.ts.map +1 -0
  579. package/lib/previewer/artifacts.js +180 -0
  580. package/lib/previewer/artifacts.js.map +1 -0
  581. package/lib/previewer/server.d.ts +111 -0
  582. package/lib/previewer/server.d.ts.map +1 -0
  583. package/lib/previewer/server.js +559 -0
  584. package/lib/previewer/server.js.map +1 -0
  585. package/package.json +4 -2
  586. package/skills/lux-answerer/SKILL.md +2 -2
  587. package/src/answerers/claude-code/index.ts +6 -5
  588. package/src/answerers/persona/index.ts +80 -34
  589. package/src/answerers/terminal/index.ts +7 -11
  590. package/src/assertions/authoring.ts +2 -0
  591. package/src/assertions/core/command-exits-zero.ts +82 -17
  592. package/src/assertions/core/event-checks.ts +64 -3
  593. package/src/assertions/core/index.ts +2 -0
  594. package/src/assertions/core/resource-checks.ts +38 -5
  595. package/src/assertions/core/safe-regex.ts +16 -6
  596. package/src/assertions/rubric/evidence.ts +85 -0
  597. package/src/assertions/rubric/index.ts +76 -46
  598. package/src/assertions/rubric/internal.ts +264 -46
  599. package/src/authoring.ts +346 -2
  600. package/src/cli/bin.ts +2 -12
  601. package/src/cli/command-runtime.ts +20 -9
  602. package/src/cli/doctor.ts +17 -0
  603. package/src/cli/harness-plan.ts +123 -0
  604. package/src/cli/init-templates.ts +176 -0
  605. package/src/cli/init.ts +53 -184
  606. package/src/cli/output-schemas.ts +109 -0
  607. package/src/cli/program.ts +697 -730
  608. package/src/cli/render/format.ts +1 -1
  609. package/src/cli/render/reporter.ts +36 -8
  610. package/src/cli/run-options.ts +131 -25
  611. package/src/cli/view.ts +129 -0
  612. package/src/container/driver.ts +213 -0
  613. package/src/container/paths.ts +29 -0
  614. package/src/container/protocol.ts +48 -0
  615. package/src/container/resources.ts +165 -0
  616. package/src/container/worker.ts +107 -0
  617. package/src/core/action-events.ts +59 -0
  618. package/src/core/agent-activity.ts +236 -0
  619. package/src/core/agent-environment.ts +64 -0
  620. package/src/core/agent-events.ts +17 -0
  621. package/src/core/annotation.ts +2 -1
  622. package/src/core/answerer.ts +9 -0
  623. package/src/core/artifact-evidence.ts +40 -5
  624. package/src/core/assertion.ts +8 -2
  625. package/src/core/case-defaults.ts +65 -0
  626. package/src/core/case.ts +79 -8
  627. package/src/core/command-verification.ts +74 -0
  628. package/src/core/driver-execution.ts +84 -0
  629. package/src/core/driver-metrics.ts +51 -0
  630. package/src/core/driver.ts +76 -0
  631. package/src/core/duration.ts +52 -0
  632. package/src/core/environment.ts +22 -0
  633. package/src/core/evaluation-clusters.ts +35 -0
  634. package/src/core/events.ts +13 -2
  635. package/src/core/experiment.ts +71 -25
  636. package/src/core/file-tree.ts +35 -0
  637. package/src/core/global-id.ts +26 -0
  638. package/src/core/harness-catalog.ts +100 -0
  639. package/src/core/harness-catalogs/antigravity.ts +24 -0
  640. package/src/core/harness-catalogs/claude-code.ts +25 -0
  641. package/src/core/harness-catalogs/codex.ts +27 -0
  642. package/src/core/harness-catalogs/copilot.ts +46 -0
  643. package/src/core/harness-catalogs/cursor.ts +13 -0
  644. package/src/core/harness-catalogs/grok-build.ts +19 -0
  645. package/src/core/harness-catalogs/index.ts +31 -0
  646. package/src/core/harness-matrix.ts +316 -0
  647. package/src/core/index.ts +12 -0
  648. package/src/core/lifecycle-fixtures.ts +4 -2
  649. package/src/core/model-cli.ts +78 -28
  650. package/src/core/platform-process.ts +65 -16
  651. package/src/core/reference-defaults.ts +47 -0
  652. package/src/core/run.ts +31 -0
  653. package/src/core/usage.ts +8 -0
  654. package/src/drivers/antigravity/README.md +171 -0
  655. package/src/drivers/antigravity/artifacts.ts +37 -0
  656. package/src/drivers/antigravity/config.ts +92 -0
  657. package/src/drivers/antigravity/events.ts +290 -0
  658. package/src/drivers/antigravity/index.ts +132 -0
  659. package/src/drivers/antigravity/runtime.ts +94 -0
  660. package/src/drivers/claude-code/deferred-process.ts +118 -0
  661. package/src/drivers/claude-code/index.ts +228 -282
  662. package/src/drivers/claude-code/isolation.ts +228 -0
  663. package/src/drivers/claude-code/parse-events.ts +98 -12
  664. package/src/drivers/codex/app-events.ts +126 -16
  665. package/src/drivers/codex/config.ts +59 -8
  666. package/src/drivers/codex/exec-events.ts +258 -0
  667. package/src/drivers/codex/index.ts +146 -312
  668. package/src/drivers/copilot/README.md +60 -0
  669. package/src/drivers/copilot/config.ts +61 -0
  670. package/src/drivers/copilot/index.ts +310 -0
  671. package/src/drivers/cursor/README.md +66 -0
  672. package/src/drivers/cursor/config.ts +31 -0
  673. package/src/drivers/cursor/index.ts +326 -0
  674. package/src/drivers/cursor/interaction.ts +74 -0
  675. package/src/drivers/grok-build/README.md +112 -0
  676. package/src/drivers/grok-build/config.ts +43 -0
  677. package/src/drivers/grok-build/index.ts +331 -0
  678. package/src/drivers/mcp/index.ts +25 -11
  679. package/src/drivers/shared/acp-events.ts +94 -0
  680. package/src/drivers/shared/acp-interaction.ts +27 -0
  681. package/src/drivers/shared/acp-session.ts +40 -0
  682. package/src/drivers/shared/acp-transport.ts +187 -0
  683. package/src/drivers/shared/artifacts.ts +6 -2
  684. package/src/drivers/shared/experiment-behavior.ts +217 -0
  685. package/src/drivers/shared/experiment-policy.ts +201 -0
  686. package/src/drivers/shared/skills.ts +74 -0
  687. package/src/drivers/subprocess/index.ts +55 -12
  688. package/src/index.ts +117 -6
  689. package/src/optimization/reflective.ts +3 -2
  690. package/src/orchestrator/annotation-loader.ts +10 -0
  691. package/src/orchestrator/annotation-store.ts +24 -3
  692. package/src/orchestrator/artifact-origin.ts +50 -0
  693. package/src/orchestrator/campaign-lifecycle.ts +5 -1
  694. package/src/orchestrator/candidates.ts +3 -2
  695. package/src/orchestrator/compare.ts +62 -13
  696. package/src/orchestrator/comparison-identity.ts +2 -0
  697. package/src/orchestrator/config.ts +23 -1
  698. package/src/orchestrator/discover.ts +8 -1
  699. package/src/orchestrator/doctor.ts +281 -74
  700. package/src/orchestrator/driver-capabilities.ts +21 -0
  701. package/src/orchestrator/driver-identity.ts +69 -0
  702. package/src/orchestrator/experiment-context.ts +5 -3
  703. package/src/orchestrator/experiment-corpus.ts +68 -39
  704. package/src/orchestrator/experiment-effects.ts +63 -0
  705. package/src/orchestrator/experiment-evaluation.ts +88 -10
  706. package/src/orchestrator/experiment-identity.ts +98 -61
  707. package/src/orchestrator/experiment-preflight.ts +27 -0
  708. package/src/orchestrator/experiment-report.ts +21 -1
  709. package/src/orchestrator/experiment-runs.ts +11 -4
  710. package/src/orchestrator/experiment-runtime-identity.ts +5 -243
  711. package/src/orchestrator/experiment-selection.ts +178 -18
  712. package/src/orchestrator/experiment.ts +29 -29
  713. package/src/orchestrator/fixture-identity.ts +9 -1
  714. package/src/orchestrator/fs-read.ts +2 -32
  715. package/src/orchestrator/grader-experiment.ts +74 -13
  716. package/src/orchestrator/grading.ts +60 -31
  717. package/src/orchestrator/harness-catalog-drift.ts +132 -0
  718. package/src/orchestrator/harness-registry.ts +12 -0
  719. package/src/orchestrator/index.ts +1 -0
  720. package/src/orchestrator/list-runs.ts +102 -18
  721. package/src/orchestrator/optimization-lifecycle.ts +4 -1
  722. package/src/orchestrator/orchestrator.ts +104 -226
  723. package/src/orchestrator/paired-effects.ts +214 -0
  724. package/src/orchestrator/profile.ts +24 -0
  725. package/src/orchestrator/provenance.ts +23 -3
  726. package/src/orchestrator/regrade.ts +5 -1
  727. package/src/orchestrator/rubric-loader.ts +1 -1
  728. package/src/orchestrator/run-dir.ts +8 -1
  729. package/src/orchestrator/run-execution.ts +75 -10
  730. package/src/orchestrator/run-lifecycle.ts +4 -2
  731. package/src/orchestrator/run-workspace.ts +62 -0
  732. package/src/orchestrator/suite-compare.ts +47 -23
  733. package/src/orchestrator/suite-runs.ts +13 -2
  734. package/src/orchestrator/suite-summary.ts +129 -15
  735. package/src/previewer/artifacts.ts +221 -0
  736. package/src/previewer/server.ts +789 -0
  737. package/viewer/app.css +2111 -0
  738. package/viewer/app.js +627 -0
  739. package/viewer/artifact-view.js +48 -0
  740. package/viewer/charts.js +98 -0
  741. package/viewer/detail.js +1266 -0
  742. package/viewer/format.js +129 -0
  743. package/viewer/index.html +151 -0
  744. package/lib/cli/skills.d.ts +0 -38
  745. package/lib/cli/skills.d.ts.map +0 -1
  746. package/lib/cli/skills.js +0 -84
  747. package/lib/cli/skills.js.map +0 -1
  748. package/lib/drivers/shared/secure-copy.d.ts.map +0 -1
  749. package/lib/drivers/shared/secure-copy.js.map +0 -1
  750. package/skills/lux/SKILL.md +0 -89
  751. package/skills/lux/references/cli.md +0 -125
  752. package/skills/lux/references/eval-authoring.md +0 -257
  753. package/src/cli/skills.ts +0 -129
  754. /package/lib/{drivers/shared → core}/secure-copy.d.ts +0 -0
  755. /package/lib/{drivers/shared → core}/secure-copy.js +0 -0
  756. /package/src/{drivers/shared → core}/secure-copy.ts +0 -0
@@ -0,0 +1,1266 @@
1
+ const assertionScore = (assertion) =>
2
+ assertion.errorMessage !== undefined || assertion.abstention !== undefined
3
+ ? null
4
+ : assertion.score;
5
+
6
+ const assertionPresentation = (assertion) => {
7
+ if (assertionScore(assertion) === null) return { className: " unavailable", mark: "?", tone: "" };
8
+ if (assertion.passed === null) return { className: " informational", mark: "ℹ", tone: "" };
9
+ if (assertion.passed) return { className: " passed", mark: "✓", tone: "tone-positive" };
10
+ return { className: "", mark: "×", tone: "tone-negative" };
11
+ };
12
+
13
+ import { comparisonBars, compositionFigure, quantizedBar, scoreBar } from "./charts.js";
14
+ import {
15
+ badge,
16
+ compactNumber,
17
+ detailEmpty,
18
+ duration,
19
+ escapeHtml,
20
+ finite,
21
+ fullDate,
22
+ metricName,
23
+ metricValue,
24
+ percentage,
25
+ signedDuration,
26
+ signedPercentage,
27
+ statusLabel,
28
+ toneForDelta,
29
+ totalTokens,
30
+ } from "./format.js";
31
+
32
+ const json = (value) => {
33
+ if (value === undefined) return "";
34
+ try {
35
+ return JSON.stringify(value, null, 2);
36
+ } catch {
37
+ return String(value);
38
+ }
39
+ };
40
+
41
+ const eventContent = (event) => {
42
+ if (event.type === "lifecycle") return [event.stage, "Run lifecycle"];
43
+ if (event.type === "ready") {
44
+ return [
45
+ `Agent ready · ${event.agentId}`,
46
+ event.model ? `Model ${event.model}` : "Driver ready",
47
+ ];
48
+ }
49
+ if (event.type === "interrupt") {
50
+ return [
51
+ `Human-in-the-loop · ${event.kind}`,
52
+ event.question ?? event.request ?? json(event.payload),
53
+ ];
54
+ }
55
+ if (event.type === "responded") {
56
+ let answer = event.text ?? null;
57
+ if (!answer && typeof event.approved === "boolean") {
58
+ answer = event.approved ? "Approved" : "Denied";
59
+ }
60
+ return [`Answerer · ${event.kind}`, answer ?? json(event.payload)];
61
+ }
62
+ if (event.type === "progress") {
63
+ const payload = event.payload ?? {};
64
+ const actor = event.agent?.role === "subagent" ? ` · Subagent ${event.agent.id}` : "";
65
+ if (event.kind === "agent-spawn") {
66
+ const spawned = payload.agent ?? {};
67
+ return [
68
+ `Subagent spawned · ${spawned.name ?? spawned.id ?? "unknown"}`,
69
+ [spawned.id, spawned.model, payload.prompt].filter(Boolean).join(" · "),
70
+ ];
71
+ }
72
+ if (event.kind === "agent-completed") {
73
+ return [
74
+ `Subagent ${payload.status ?? "completed"} · ${payload.agentId ?? "unknown"}`,
75
+ "Agent lifecycle",
76
+ ];
77
+ }
78
+ if (event.kind === "action") {
79
+ return [
80
+ `Action${actor} · ${payload.name ?? "unknown"}`,
81
+ payload.category ? `Category: ${payload.category}` : json(payload),
82
+ ];
83
+ }
84
+ const name = event.kind === "tool-call" ? payload.name : event.kind;
85
+ return [`Progress${actor} · ${name ?? event.kind}`, json(payload)];
86
+ }
87
+ if (event.type === "error") return [`Error · ${event.exitReason}`, event.reason];
88
+ return ["Completed", event.summary ? json(event.summary) : "Driver finished"];
89
+ };
90
+
91
+ const eventTime = (event, startedAt) => {
92
+ if (typeof event.ts !== "number") return "setup";
93
+ if (event.ts > 1_000_000_000_000) return fullDate(event.ts);
94
+ const start = Date.parse(startedAt ?? "");
95
+ if (!Number.isFinite(start)) return duration(event.ts);
96
+ return `+${duration(Math.max(0, event.ts - start))}`;
97
+ };
98
+
99
+ const assertionScoreLabel = (assertion) => {
100
+ const label = `${assertion.name} scored ${percentage(assertionScore(assertion))}`;
101
+ if (assertion.informational || assertion.passed === null) return `${label} (informational)`;
102
+ if (typeof assertion.threshold === "number") {
103
+ return `${label}; threshold ${percentage(assertion.threshold)}`;
104
+ }
105
+ return label;
106
+ };
107
+
108
+ const assertionMeta = (assertion) =>
109
+ [
110
+ assertion.type,
111
+ !assertion.informational && assertion.passed !== null && typeof assertion.threshold === "number"
112
+ ? `threshold ${finite(assertion.threshold)}`
113
+ : null,
114
+ `weight ${finite(assertion.weight)}`,
115
+ duration(assertion.durationMs),
116
+ assertion.usage ? `${compactNumber(totalTokens(assertion.usage))} judge tokens` : null,
117
+ ]
118
+ .filter(Boolean)
119
+ .join(" · ");
120
+
121
+ const assertionDiagnostics = (assertion) => {
122
+ const rows = [
123
+ assertion.comment ? `<p class="assertion-comment">${escapeHtml(assertion.comment)}</p>` : "",
124
+ assertion.abstention ? `<p class="assertion-comment">Insufficient evidence</p>` : "",
125
+ assertion.errorMessage
126
+ ? `<p class="diagnostic-error">${escapeHtml(assertion.errorMessage)}</p>`
127
+ : "",
128
+ assertion.detail !== undefined
129
+ ? `<details class="diagnostic"><summary>Structured detail</summary><pre>${escapeHtml(
130
+ json(assertion.detail),
131
+ )}</pre></details>`
132
+ : "",
133
+ ];
134
+ return rows.join("");
135
+ };
136
+
137
+ const inferredMetricUnit = (name, unit) => {
138
+ if (unit) return unit;
139
+ if (name === "score") return "score";
140
+ if (name.endsWith("Ms")) return "milliseconds";
141
+ if (name.endsWith(".bytes")) return "bytes";
142
+ if (name.endsWith(".usd")) return "usd";
143
+ if (name.endsWith("Rate")) return "ratio";
144
+ if (
145
+ name.includes("tokens") ||
146
+ /(?:^|\.)(?:input|output|total|cacheRead|cacheCreation)$/.test(name)
147
+ ) {
148
+ return "tokens";
149
+ }
150
+ return "number";
151
+ };
152
+
153
+ const driverMetricStats = (metrics) =>
154
+ Object.entries(metrics ?? {})
155
+ .map(
156
+ ([name, metric]) =>
157
+ `<div><dt>${escapeHtml(metricName(name, metric.label))}</dt><dd>${escapeHtml(
158
+ metricValue(metric.value, metric.unit),
159
+ )}</dd><small>${escapeHtml(name)}</small></div>`,
160
+ )
161
+ .join("");
162
+
163
+ const agentActivityCards = (activities) =>
164
+ (activities ?? [])
165
+ .filter((activity) => activity.agent?.role === "subagent")
166
+ .map((activity, index) => {
167
+ const agent = activity.agent;
168
+ const actions = activity.actions ?? {};
169
+ const tools = activity.toolCalls ?? {};
170
+ const label = agent.name ?? `Subagent ${index + 1}`;
171
+ return `<article class="agent-card">
172
+ <header><div><strong>${escapeHtml(label)}</strong><small>${escapeHtml(agent.id)}</small></div><span>${escapeHtml(
173
+ `${activity.status ?? "unknown"} · ${duration(activity.durationMs)}`,
174
+ )}</span></header>
175
+ <dl>
176
+ <div><dt>Events</dt><dd>${compactNumber(activity.eventCount)}</dd></div>
177
+ <div><dt>Tokens</dt><dd>${compactNumber(activity.usage?.total)}</dd></div>
178
+ <div><dt>Tool calls</dt><dd>${compactNumber(tools.total)}</dd></div>
179
+ <div><dt>Succeeded</dt><dd>${compactNumber(tools.succeeded)}</dd></div>
180
+ <div><dt>Failed</dt><dd>${compactNumber(tools.failed)}</dd></div>
181
+ <div><dt>Actions</dt><dd>${compactNumber(actions.total)}</dd></div>
182
+ <div><dt>Mutations</dt><dd>${compactNumber(actions.mutations)}</dd></div>
183
+ <div><dt>First action</dt><dd>${escapeHtml(duration(actions.timeToFirstActionMs))}</dd></div>
184
+ <div><dt>First mutation</dt><dd>${escapeHtml(duration(actions.timeToFirstMutationMs))}</dd></div>
185
+ </dl>
186
+ <small class="agent-parent">Spawn ${escapeHtml(agent.spawnToolUseId ?? "unknown")} · parent ${escapeHtml(
187
+ agent.parentId ?? "unknown",
188
+ )}</small>
189
+ </article>`;
190
+ })
191
+ .join("");
192
+
193
+ export const renderRunDetail = (item, detail) => {
194
+ const metadata = detail.metadata;
195
+ const grading = detail.grading;
196
+ const assertions = grading?.assertions ?? [];
197
+ const events = detail.events ?? [];
198
+ const artifacts = detail.artifacts ?? [];
199
+ const reportedAgentTokens = totalTokens(metadata?.usage?.agent);
200
+ const agentTokens =
201
+ metadata?.agentUsageAvailability === "reported" ||
202
+ (metadata?.agentUsageAvailability === undefined && reportedAgentTokens > 0)
203
+ ? reportedAgentTokens
204
+ : undefined;
205
+ const answererTokens = totalTokens(metadata?.usage?.answerer);
206
+ const gradingTokens = totalTokens(metadata?.usage?.grading);
207
+ const harnessTokens = answererTokens + gradingTokens;
208
+ const phase = metadata?.phaseDurationMs;
209
+ const actionTiming = metadata?.actionTiming;
210
+ const prompt = detail.case?.prompt;
211
+ const conditions = metadata?.provenance?.environment;
212
+ const host = metadata?.provenance?.runtime?.host;
213
+ const driverMetrics = driverMetricStats(metadata?.driverMetrics);
214
+ const subagents = agentActivityCards(metadata?.agentActivity);
215
+ return `
216
+ <header class="detail-header">
217
+ <div class="detail-kicker">
218
+ ${badge(item.status)}
219
+ <span class="detail-path">${escapeHtml(item.path)}</span>
220
+ </div>
221
+ <div class="detail-title-row">
222
+ <div>
223
+ <h2>${escapeHtml(detail.case?.id ?? item.caseId)}</h2>
224
+ <p class="detail-subtitle">${escapeHtml(
225
+ [metadata?.driver?.name, metadata?.driver?.model, fullDate(metadata?.startedAt)]
226
+ .filter(Boolean)
227
+ .join(" · "),
228
+ )}</p>
229
+ </div>
230
+ <div class="score-block">
231
+ <strong>${escapeHtml(percentage(grading?.caseScore))}</strong>
232
+ <small>case score</small>
233
+ </div>
234
+ </div>
235
+ </header>
236
+ <div class="detail-body">
237
+ ${
238
+ grading?.incomplete
239
+ ? `<div class="attention-banner">Incomplete ${escapeHtml(
240
+ grading.incomplete.phase ?? "run",
241
+ )}: ${escapeHtml(grading.incomplete.exitReason)}</div>`
242
+ : ""
243
+ }
244
+ <dl class="detail-stats">
245
+ <div><dt>Duration</dt><dd>${escapeHtml(duration(item.durationMs))}</dd></div>
246
+ <div><dt>Agent tokens</dt><dd>${escapeHtml(compactNumber(agentTokens))}</dd></div>
247
+ <div><dt>Harness tokens</dt><dd>${escapeHtml(compactNumber(harnessTokens))}</dd></div>
248
+ <div><dt>Artifacts</dt><dd>${artifacts.length}</dd></div>
249
+ <div><dt>First action</dt><dd>${escapeHtml(duration(actionTiming?.timeToFirstActionMs))}</dd></div>
250
+ <div><dt>First mutation</dt><dd>${escapeHtml(duration(actionTiming?.timeToFirstMutationMs))}</dd></div>
251
+ <div><dt>First durable mutation</dt><dd>${escapeHtml(
252
+ duration(actionTiming?.timeToFirstDurableMutationMs),
253
+ )}</dd></div>
254
+ </dl>
255
+ ${
256
+ conditions
257
+ ? `<section class="detail-section"><h3>Execution conditions</h3>
258
+ <p>Lux concurrency: ${escapeHtml(conditions.concurrency)}${host ? ` · Host: ${escapeHtml(host.availableParallelism)} available CPUs, ${escapeHtml(metricValue(host.totalMemoryBytes, "bytes"))} memory` : ""}</p>
259
+ ${conditions.declared ? `<details><summary>Declared environment (not enforced by Lux)</summary><pre>${escapeHtml(JSON.stringify(conditions.declared, null, 2))}</pre></details>` : ""}
260
+ </section>`
261
+ : ""
262
+ }
263
+ ${
264
+ driverMetrics
265
+ ? `<section class="detail-section"><h3>Driver metrics <span>reported by ${escapeHtml(
266
+ metadata?.driver?.name ?? "driver",
267
+ )}</span></h3><dl class="detail-stats driver-metrics">${driverMetrics}</dl></section>`
268
+ : ""
269
+ }
270
+ ${
271
+ subagents
272
+ ? `<section class="detail-section"><h3>Subagent activity <span>${
273
+ metadata.agentActivity.filter((activity) => activity.agent.role === "subagent").length
274
+ } spawned</span></h3><div class="agent-grid">${subagents}</div></section>`
275
+ : ""
276
+ }
277
+ <div class="evidence-visual-grid">
278
+ ${compositionFigure(
279
+ "Judgment outcome",
280
+ [
281
+ {
282
+ key: "passed",
283
+ label: "Passed",
284
+ value: assertions.filter((assertion) => assertion.passed).length,
285
+ },
286
+ {
287
+ key: "regressed",
288
+ label: "Failed",
289
+ value: assertions.filter((assertion) => assertion.passed === false).length,
290
+ },
291
+ ],
292
+ `${assertions.filter((assertion) => assertion.passed).length} of ${
293
+ assertions.filter((assertion) => assertion.passed !== null).length
294
+ } assertions passed`,
295
+ )}
296
+ ${
297
+ agentTokens + answererTokens + gradingTokens > 0
298
+ ? compositionFigure(
299
+ "Token composition",
300
+ [
301
+ { key: "agent", label: "Agent", value: agentTokens },
302
+ { key: "answerer", label: "Answerer", value: answererTokens },
303
+ { key: "grading", label: "Grading", value: gradingTokens },
304
+ ],
305
+ `${agentTokens ?? "unavailable"} agent, ${answererTokens} answerer, ${gradingTokens} grading tokens`,
306
+ )
307
+ : ""
308
+ }
309
+ ${
310
+ phase
311
+ ? compositionFigure(
312
+ "Runtime composition",
313
+ [
314
+ { key: "agent", label: "Agent", value: phase.agent },
315
+ { key: "collect", label: "Collect", value: phase.collect },
316
+ { key: "grading", label: "Grading", value: phase.grading },
317
+ ],
318
+ `${duration(phase.agent)} agent, ${duration(phase.collect)} collection, ${duration(
319
+ phase.grading,
320
+ )} grading`,
321
+ )
322
+ : ""
323
+ }
324
+ </div>
325
+ ${
326
+ prompt
327
+ ? `<section class="detail-section"><h3>Case prompt</h3><div class="prompt-block">${escapeHtml(
328
+ prompt,
329
+ )}</div></section>`
330
+ : ""
331
+ }
332
+ <section class="detail-section">
333
+ <h3>Judgment <span>${assertions.filter((entry) => entry.passed).length}/${assertions.filter((entry) => entry.passed !== null).length} passed</span></h3>
334
+ <div class="assertion-list">
335
+ ${
336
+ assertions.length
337
+ ? assertions
338
+ .map(
339
+ (assertion) => `
340
+ <article class="assertion${assertionPresentation(assertion).className}">
341
+ <span class="assertion-mark">${assertionPresentation(assertion).mark}</span>
342
+ <span class="assertion-content">
343
+ <strong>${escapeHtml(assertion.name)}</strong>
344
+ <small>${escapeHtml(assertionMeta(assertion))}</small>
345
+ ${scoreBar(
346
+ assertionScore(assertion),
347
+ assertionPresentation(assertion).tone,
348
+ assertionScoreLabel(assertion),
349
+ assertion.informational || assertion.passed === null
350
+ ? null
351
+ : assertion.threshold,
352
+ )}
353
+ ${assertionDiagnostics(assertion)}
354
+ </span>
355
+ <span class="assertion-score">${escapeHtml(
356
+ percentage(assertionScore(assertion)),
357
+ )}</span>
358
+ </article>`,
359
+ )
360
+ .join("")
361
+ : '<div class="artifact"><span>No grading assertions recorded</span><small>ungraded</small></div>'
362
+ }
363
+ </div>
364
+ </section>
365
+ <section class="detail-section">
366
+ <h3>Trace <span>${events.length} events</span></h3>
367
+ <div class="trace-list">
368
+ ${
369
+ events.length
370
+ ? events
371
+ .map((event) => {
372
+ const [title, content] = eventContent(event);
373
+ return `<article class="trace-item">
374
+ <time>${escapeHtml(eventTime(event, metadata?.startedAt))}</time>
375
+ <div><strong>${escapeHtml(title)}</strong><p>${escapeHtml(content)}</p></div>
376
+ </article>`;
377
+ })
378
+ .join("")
379
+ : '<div class="artifact"><span>No trace events recorded</span><small>empty</small></div>'
380
+ }
381
+ </div>
382
+ </section>
383
+ <section class="detail-section">
384
+ <h3>Artifacts <span>${artifacts.length} files</span></h3>
385
+ <div class="artifact-list">
386
+ ${
387
+ artifacts.length
388
+ ? artifacts
389
+ .map(
390
+ (artifact) => `<button class="artifact artifact-button" type="button"
391
+ data-artifact="${escapeHtml(artifact)}">
392
+ <span>${escapeHtml(artifact)}</span><small>preview →</small>
393
+ </button>`,
394
+ )
395
+ .join("")
396
+ : '<div class="artifact"><span>No captured artifacts</span><small>empty</small></div>'
397
+ }
398
+ </div>
399
+ </section>
400
+ </div>`;
401
+ };
402
+
403
+ const suiteItemStatus = (item) => {
404
+ if (item.status !== "completed") return item.status;
405
+ if (item.exitReason !== "done" || item.casePassed === false) return "failed";
406
+ if (item.casePassed === true) return "passed";
407
+ return "ungraded";
408
+ };
409
+
410
+ const metricSummaryRows = (metrics) =>
411
+ Object.entries(metrics ?? {})
412
+ .map(
413
+ ([name, summary]) => `<tr>
414
+ <td>${escapeHtml(name)}</td>
415
+ <td>${escapeHtml(metricValue(summary.mean, inferredMetricUnit(name, summary.unit)))}</td>
416
+ <td>${escapeHtml(metricValue(summary.median, inferredMetricUnit(name, summary.unit)))}</td>
417
+ <td>${escapeHtml(metricValue(summary.p90, inferredMetricUnit(name, summary.unit)))}</td>
418
+ <td>${summary.n}</td>
419
+ </tr>`,
420
+ )
421
+ .join("");
422
+
423
+ export const renderSuiteDetail = (item, payload) => {
424
+ const summary = payload?.summary;
425
+ const totals = summary?.totals;
426
+ const metrics = metricSummaryRows(summary?.metrics);
427
+ return `
428
+ <header class="detail-header">
429
+ <div class="detail-kicker">${badge(item.status)}<span class="detail-path">${escapeHtml(
430
+ item.path,
431
+ )}</span></div>
432
+ <div class="detail-title-row">
433
+ <div><h2>${escapeHtml(item.suiteId)}</h2><p class="detail-subtitle">${escapeHtml(
434
+ `${item.profileId ?? "default profile"} · ${fullDate(item.startedAt)}`,
435
+ )}</p></div>
436
+ <div class="score-block"><strong>${escapeHtml(
437
+ percentage(totals?.passRate),
438
+ )}</strong><small>${totals?.passed ?? 0}/${totals?.attempts ?? item.items.length} cases passed</small></div>
439
+ </div>
440
+ </header>
441
+ <div class="detail-body">
442
+ <dl class="detail-stats">
443
+ <div><dt>Attempts</dt><dd>${totals?.attempts ?? item.items.length}</dd></div>
444
+ <div><dt>Concurrency</dt><dd>${item.concurrency}</dd></div>
445
+ <div><dt>Failed</dt><dd>${totals?.failed ?? 0}</dd></div>
446
+ <div><dt>Pending</dt><dd>${totals?.pending ?? 0}</dd></div>
447
+ </dl>
448
+ ${compositionFigure(
449
+ "Suite outcomes",
450
+ [
451
+ { key: "passed", label: "Passed", value: totals?.passed ?? 0 },
452
+ { key: "regressed", label: "Failed", value: totals?.failed ?? 0 },
453
+ { key: "ungraded", label: "Ungraded", value: totals?.ungraded ?? 0 },
454
+ { key: "pending", label: "Pending", value: totals?.pending ?? 0 },
455
+ ],
456
+ `${totals?.passed ?? 0} passed, ${totals?.failed ?? 0} failed, ${
457
+ totals?.ungraded ?? 0
458
+ } ungraded, ${totals?.pending ?? 0} pending`,
459
+ )}
460
+ ${
461
+ metrics
462
+ ? `<section class="detail-section">
463
+ <h3>Aggregate metrics <span>mean / median / p90</span></h3>
464
+ <div class="matrix-wrap"><table class="matrix">
465
+ <thead><tr><th>Metric</th><th>Mean</th><th>Median</th><th>P90</th><th>N</th></tr></thead>
466
+ <tbody>${metrics}</tbody>
467
+ </table></div>
468
+ </section>`
469
+ : ""
470
+ }
471
+ <section class="detail-section">
472
+ <h3>Case attempts <span>${item.items.length} total</span></h3>
473
+ <div class="suite-list">
474
+ ${item.items
475
+ .map((entry) => {
476
+ const content = `<span>${escapeHtml(entry.caseId)}</span>
477
+ <small>${escapeHtml(statusLabel(suiteItemStatus(entry)))}</small>`;
478
+ return entry.runId
479
+ ? `<button class="suite-item suite-button" type="button" data-open-run="${escapeHtml(
480
+ entry.runId,
481
+ )}">${content}</button>`
482
+ : `<div class="suite-item">${content}</div>`;
483
+ })
484
+ .join("")}
485
+ </div>
486
+ </section>
487
+ </div>`;
488
+ };
489
+
490
+ const candidateById = (payload, id) => payload.candidates.find((candidate) => candidate.id === id);
491
+
492
+ const baselineCandidate = (payload) =>
493
+ candidateById(payload, "baseline") ??
494
+ payload.candidates.find((candidate) => candidate.parentIds.length === 0) ??
495
+ payload.candidates[0] ??
496
+ null;
497
+
498
+ const evaluationKey = (evaluation) => `${evaluation.profileId}/${evaluation.split}`;
499
+
500
+ const caseDeltaRows = (baseline, target) => {
501
+ const cases = new Set([
502
+ ...Object.keys(baseline?.casePassRates ?? {}),
503
+ ...Object.keys(target?.casePassRates ?? {}),
504
+ ]);
505
+ return [...cases]
506
+ .sort()
507
+ .map((caseId) => {
508
+ const before = baseline?.casePassRates?.[caseId];
509
+ const after = target?.casePassRates?.[caseId];
510
+ const delta = typeof before === "number" && typeof after === "number" ? after - before : null;
511
+ return `<tr>
512
+ <td>${escapeHtml(caseId)}</td>
513
+ <td class="bar-table-cell">${quantizedBar(
514
+ before,
515
+ "tone-before",
516
+ `${caseId} baseline pass rate ${percentage(before)}`,
517
+ )}<small>${escapeHtml(percentage(before))}</small></td>
518
+ <td class="bar-table-cell">${quantizedBar(
519
+ after,
520
+ "tone-after",
521
+ `${caseId} candidate pass rate ${percentage(after)}`,
522
+ )}<small>${escapeHtml(percentage(after))}</small></td>
523
+ <td class="${toneForDelta(delta)}">${escapeHtml(signedPercentage(delta))}</td>
524
+ </tr>`;
525
+ })
526
+ .join("");
527
+ };
528
+
529
+ const evaluationMetricComparisons = (before, after) => {
530
+ const definitions = { ...(before?.metricDefinitions ?? {}), ...(after?.metricDefinitions ?? {}) };
531
+ return Object.keys(definitions)
532
+ .sort()
533
+ .map((name) => {
534
+ const definition = definitions[name];
535
+ const a = before?.metrics?.[name];
536
+ const b = after?.metrics?.[name];
537
+ const delta = typeof a === "number" && typeof b === "number" ? b - a : null;
538
+ const tone =
539
+ definition.better === "neutral"
540
+ ? "metric-neutral"
541
+ : toneForDelta(delta, definition.better === "lower");
542
+ return comparisonBars({
543
+ title: metricName(name, definition.label),
544
+ before: a,
545
+ after: b,
546
+ beforeText: metricValue(a, definition.unit),
547
+ afterText: metricValue(b, definition.unit),
548
+ deltaText: signedMetricValue(delta, definition.unit),
549
+ deltaTone: tone,
550
+ });
551
+ })
552
+ .join("");
553
+ };
554
+
555
+ const experimentScopes = (baselineEvaluations, targetEvaluations) => {
556
+ const baselineByKey = new Map(baselineEvaluations.map((entry) => [evaluationKey(entry), entry]));
557
+ const targetByKey = new Map(targetEvaluations.map((entry) => [evaluationKey(entry), entry]));
558
+ const keys = new Set([...baselineByKey.keys(), ...targetByKey.keys()]);
559
+ return [...keys]
560
+ .sort()
561
+ .map((key) => {
562
+ const before = baselineByKey.get(key);
563
+ const after = targetByKey.get(key);
564
+ const delta =
565
+ typeof before?.passRate === "number" && typeof after?.passRate === "number"
566
+ ? after.passRate - before.passRate
567
+ : null;
568
+ const metrics = evaluationMetricComparisons(before, after);
569
+ return `<article class="comparison-scope">
570
+ <header>
571
+ <strong>${escapeHtml(key)}</strong>
572
+ <span class="${toneForDelta(delta)}">${escapeHtml(signedPercentage(delta))}</span>
573
+ </header>
574
+ <div class="matrix-wrap"><table class="matrix">
575
+ <thead><tr><th>Case</th><th>Baseline</th><th>Candidate</th><th>Δ</th></tr></thead>
576
+ <tbody>${caseDeltaRows(before, after) || '<tr><td colspan="4">No per-case evidence.</td></tr>'}</tbody>
577
+ </table></div>
578
+ ${metrics ? `<div class="comparison-plot-grid experiment-metrics">${metrics}</div>` : ""}
579
+ </article>`;
580
+ })
581
+ .join("");
582
+ };
583
+
584
+ const candidateLandscape = (evaluations, baselineId, targetId) => {
585
+ const rows = evaluations
586
+ .slice()
587
+ .sort((left, right) =>
588
+ `${left.profileId}/${left.split}/${left.candidateId}`.localeCompare(
589
+ `${right.profileId}/${right.split}/${right.candidateId}`,
590
+ ),
591
+ )
592
+ .map((evaluation) => {
593
+ let role = "peer";
594
+ let tone = "tone-neutral";
595
+ if (evaluation.candidateId === baselineId) {
596
+ role = "baseline";
597
+ tone = "tone-before";
598
+ } else if (evaluation.candidateId === targetId) {
599
+ role = "target";
600
+ tone = "tone-after";
601
+ }
602
+ return `<div class="landscape-row ${role}">
603
+ <span><strong>${escapeHtml(evaluation.candidateId)}</strong><small>${escapeHtml(
604
+ `${evaluation.profileId} / ${evaluation.split}`,
605
+ )}</small></span>
606
+ ${quantizedBar(
607
+ evaluation.passRate,
608
+ tone,
609
+ `${evaluation.candidateId} ${evaluation.profileId} ${evaluation.split} pass rate ${percentage(
610
+ evaluation.passRate,
611
+ )}`,
612
+ )}
613
+ <b>${escapeHtml(percentage(evaluation.passRate))}</b>
614
+ </div>`;
615
+ })
616
+ .join("");
617
+ return `<figure class="candidate-landscape">
618
+ <figcaption><span>Candidate pass-rate landscape</span><small>shared 0–100% scale</small></figcaption>
619
+ <div class="landscape-axis"><span>0</span><span>50</span><span>100%</span></div>
620
+ ${rows || '<div class="artifact"><span>No completed candidate evaluations</span><small>empty</small></div>'}
621
+ </figure>`;
622
+ };
623
+
624
+ const budgetUtilization = (decision) => {
625
+ if (!decision) return "";
626
+ const entries = [
627
+ ["Candidates", decision.spend.candidates, decision.budget.maxCandidates, "count"],
628
+ ["Metric calls", decision.spend.metricCalls, decision.budget.maxMetricCalls, "count"],
629
+ ["Judge calls", decision.spend.judgeCalls, decision.budget.maxJudgeCalls, "count"],
630
+ ["Proposal calls", decision.spend.proposalCalls, decision.budget.maxProposalCalls, "count"],
631
+ ["Tokens", decision.spend.totalTokens, decision.budget.maxTokens, "compact"],
632
+ ["Cost", decision.spend.costUsd, decision.budget.maxCostUsd, "currency"],
633
+ ["Active time", decision.spend.activeTimeMs, decision.budget.maxWallTimeMs, "duration"],
634
+ ];
635
+ const formatValue = (value, kind) => {
636
+ if (kind === "compact") return compactNumber(value);
637
+ if (kind === "currency") return `$${finite(value)}`;
638
+ if (kind === "duration") return duration(value);
639
+ return finite(value, 0);
640
+ };
641
+ return `<figure class="budget-visual">
642
+ <figcaption><span>Budget utilization</span><small>observed / ceiling</small></figcaption>
643
+ ${entries
644
+ .map(([label, value, limit, kind]) => {
645
+ const ratio = limit > 0 ? value / limit : 0;
646
+ return `<div class="budget-row">
647
+ <span>${escapeHtml(label)}</span>
648
+ ${quantizedBar(
649
+ ratio,
650
+ ratio > 0.9 ? "tone-negative" : "tone-neutral",
651
+ `${label} ${formatValue(value, kind)} of ${formatValue(limit, kind)}`,
652
+ )}
653
+ <b>${escapeHtml(`${formatValue(value, kind)} / ${formatValue(limit, kind)}`)}</b>
654
+ </div>`;
655
+ })
656
+ .join("")}
657
+ </figure>`;
658
+ };
659
+
660
+ const failedEvidence = (evaluations) =>
661
+ evaluations
662
+ .flatMap((evaluation) =>
663
+ evaluation.failedCases.map((failure) => ({
664
+ ...failure,
665
+ scope: evaluationKey(evaluation),
666
+ })),
667
+ )
668
+ .map(
669
+ (failure) => `<article class="failure-card">
670
+ <header><strong>${escapeHtml(failure.caseId)}</strong><small>${escapeHtml(
671
+ failure.scope,
672
+ )} · ${escapeHtml(percentage(failure.score))}</small></header>
673
+ ${
674
+ failure.feedback.length
675
+ ? `<ul>${failure.feedback.map((entry) => `<li>${escapeHtml(entry)}</li>`).join("")}</ul>`
676
+ : "<p>No evaluator feedback recorded.</p>"
677
+ }
678
+ </article>`,
679
+ )
680
+ .join("");
681
+
682
+ export const experimentTargetId = (payload, requested) => {
683
+ const baseline = baselineCandidate(payload);
684
+ const ids = new Set(payload.experiment.evaluations.map((entry) => entry.candidateId));
685
+ if (requested && ids.has(requested)) return requested;
686
+ const champion = payload.experiment.championCandidateId;
687
+ if (champion && ids.has(champion)) return champion;
688
+ return [...ids].find((id) => id !== baseline?.id) ?? [...ids][0] ?? null;
689
+ };
690
+
691
+ export const renderExperimentDetail = (item, payload, requestedTargetId) => {
692
+ const experiment = payload.experiment;
693
+ const decision = payload.decision;
694
+ const baseline = baselineCandidate(payload);
695
+ const targetId = experimentTargetId(payload, requestedTargetId);
696
+ const target = targetId ? candidateById(payload, targetId) : null;
697
+ const baselineEvaluations = experiment.evaluations.filter(
698
+ (evaluation) => evaluation.candidateId === baseline?.id,
699
+ );
700
+ const targetEvaluations = experiment.evaluations.filter(
701
+ (evaluation) => evaluation.candidateId === targetId,
702
+ );
703
+ const options = [...new Set(experiment.evaluations.map((entry) => entry.candidateId))]
704
+ .filter((id) => id !== baseline?.id || id === targetId)
705
+ .map(
706
+ (id) =>
707
+ `<option value="${escapeHtml(id)}"${id === targetId ? " selected" : ""}>${escapeHtml(
708
+ id,
709
+ )}</option>`,
710
+ )
711
+ .join("");
712
+ const gates = (decision?.gates ?? [])
713
+ .map(
714
+ (gate) => `<article class="gate ${gate.passed ? "passed" : "failed"}">
715
+ <span>${gate.passed ? "✓" : "×"}</span>
716
+ <div><strong>${escapeHtml(gate.metric)}</strong><small>${escapeHtml(
717
+ [gate.profile, gate.split, gate.interval, gate.note].filter(Boolean).join(" · "),
718
+ )}</small></div>
719
+ <b>${gate.values.map((value) => finite(value)).join(", ")}</b>
720
+ </article>`,
721
+ )
722
+ .join("");
723
+ const lineage = target
724
+ ? `<article class="lineage-card">
725
+ <span>${escapeHtml(target.id)}</span>
726
+ <small>from ${escapeHtml(target.parentIds.join(" + ") || "root")} · ${escapeHtml(
727
+ target.proposer,
728
+ )} · ${compactNumber(target.sourceBytes)} bytes</small>
729
+ ${target.rationale ? `<p>${escapeHtml(target.rationale)}</p>` : ""}
730
+ <div>${target.components
731
+ .map((component) => `<code>${escapeHtml(component)}</code>`)
732
+ .join("")}</div>
733
+ </article>`
734
+ : "";
735
+ return `
736
+ <header class="detail-header">
737
+ <div class="detail-kicker">${badge(item.status)}<span class="detail-path">${escapeHtml(
738
+ item.path,
739
+ )}</span></div>
740
+ <div class="detail-title-row">
741
+ <div><h2>${escapeHtml(item.campaignId)}</h2><p class="detail-subtitle">${escapeHtml(
742
+ `${item.mode} · ${fullDate(item.startedAt)} · ${item.experimentId}`,
743
+ )}</p></div>
744
+ <div class="score-block"><strong>${escapeHtml(
745
+ decision?.outcome ? statusLabel(decision.outcome) : item.candidateCount,
746
+ )}</strong><small>${decision?.outcome ? "verified decision" : "candidates"}</small></div>
747
+ </div>
748
+ </header>
749
+ <div class="detail-body">
750
+ ${
751
+ payload.integrity.verified
752
+ ? '<div class="integrity-banner">✓ Candidate ledger and accounting verified</div>'
753
+ : `<div class="attention-banner">Integrity unavailable: ${escapeHtml(
754
+ payload.integrity.errors.join(" · "),
755
+ )}</div>`
756
+ }
757
+ <dl class="detail-stats">
758
+ <div><dt>Champion</dt><dd>${escapeHtml(experiment.championCandidateId ?? "—")}</dd></div>
759
+ <div><dt>Active time</dt><dd>${escapeHtml(duration(experiment.activeTimeMs))}</dd></div>
760
+ <div><dt>Judge calls</dt><dd>${experiment.budgetUsage.judgeCalls}</dd></div>
761
+ <div><dt>Spend</dt><dd>$${experiment.budgetUsage.costUsd.toFixed(2)}</dd></div>
762
+ </dl>
763
+ <div class="experiment-visuals">
764
+ ${candidateLandscape(experiment.evaluations, baseline?.id, targetId)}
765
+ ${budgetUtilization(decision)}
766
+ </div>
767
+ ${
768
+ decision
769
+ ? `<section class="detail-section">
770
+ <h3>Promotion gates <span>${decision.gates.filter((gate) => gate.passed).length}/${decision.gates.length} passed</span></h3>
771
+ <div class="gate-list">${gates || '<div class="artifact"><span>No completed gates</span><small>incomplete</small></div>'}</div>
772
+ <div class="decision-meta">
773
+ <span>Ranking</span><strong>${escapeHtml(decision.ranking.join(" → ") || "—")}</strong>
774
+ <span>Eligible</span><strong>${escapeHtml(
775
+ decision.eligibleCandidateIds.join(", ") || "none",
776
+ )}</strong>
777
+ </div>
778
+ </section>`
779
+ : ""
780
+ }
781
+ <section class="detail-section">
782
+ <div class="section-heading">
783
+ <h3>Candidate comparison <span>baseline → target</span></h3>
784
+ ${
785
+ options
786
+ ? `<label class="candidate-picker">Target <select data-candidate-target>${options}</select></label>`
787
+ : ""
788
+ }
789
+ </div>
790
+ ${lineage}
791
+ <div class="comparison-scopes">
792
+ ${experimentScopes(baselineEvaluations, targetEvaluations) || '<div class="artifact"><span>No comparable evaluations</span><small>empty</small></div>'}
793
+ </div>
794
+ </section>
795
+ ${
796
+ failedEvidence(targetEvaluations)
797
+ ? `<section class="detail-section"><h3>Failed evidence <span>judge feedback</span></h3><div class="failure-list">${failedEvidence(
798
+ targetEvaluations,
799
+ )}</div></section>`
800
+ : ""
801
+ }
802
+ </div>`;
803
+ };
804
+
805
+ const detailAssertionMap = (detail) =>
806
+ new Map(
807
+ (detail.grading?.assertions ?? []).map((assertion) => [
808
+ assertion.id ?? assertion.name,
809
+ assertion,
810
+ ]),
811
+ );
812
+
813
+ const assertionDeltaRows = (payload) => {
814
+ const before = detailAssertionMap(payload.before);
815
+ const after = detailAssertionMap(payload.after);
816
+ return payload.comparison.grading.deltas
817
+ .map((delta) => {
818
+ const beforeAssertion = before.get(delta.id);
819
+ const afterAssertion = after.get(delta.id);
820
+ return `<tr>
821
+ <td><strong>${escapeHtml(delta.name)}</strong><small>${escapeHtml(delta.type)}</small></td>
822
+ <td>${escapeHtml(percentage(delta.before?.score))}</td>
823
+ <td>${escapeHtml(percentage(delta.after?.score))}</td>
824
+ <td><span class="delta-badge ${escapeHtml(delta.status)}">${escapeHtml(
825
+ delta.status,
826
+ )}</span></td>
827
+ <td class="rationale-cell">
828
+ <span>${escapeHtml(beforeAssertion?.comment ?? "—")}</span>
829
+ <b>→</b>
830
+ <span>${escapeHtml(afterAssertion?.comment ?? "—")}</span>
831
+ </td>
832
+ </tr>`;
833
+ })
834
+ .join("");
835
+ };
836
+
837
+ const artifactDelta = (payload) => {
838
+ const groups = [
839
+ ["changed", payload.comparison.artifactsDelta.changed],
840
+ ["added", payload.comparison.artifactsDelta.added],
841
+ ["removed", payload.comparison.artifactsDelta.removed],
842
+ ];
843
+ return groups
844
+ .flatMap(([status, paths]) =>
845
+ paths.map((path) => {
846
+ const clickable = status === "changed" || status === "added" || status === "removed";
847
+ const content = `<span>${escapeHtml(path)}</span><small>${escapeHtml(status)}${
848
+ clickable ? " →" : ""
849
+ }</small>`;
850
+ return clickable
851
+ ? `<button class="artifact artifact-button ${escapeHtml(
852
+ status,
853
+ )}" type="button" data-compare-artifact="${escapeHtml(path)}">${content}</button>`
854
+ : `<div class="artifact">${content}</div>`;
855
+ }),
856
+ )
857
+ .join("");
858
+ };
859
+
860
+ const signedMetricValue = (value, unit) => {
861
+ if (typeof value !== "number") return "—";
862
+ if (unit === "milliseconds") return signedDuration(value);
863
+ if (unit === "ratio" || unit === "score") return signedPercentage(value);
864
+ return `${value > 0 ? "+" : ""}${metricValue(value, unit)}`;
865
+ };
866
+
867
+ const driverMetricComparisons = (runDelta) => {
868
+ const before = runDelta?.driverMetrics?.a ?? {};
869
+ const after = runDelta?.driverMetrics?.b ?? {};
870
+ const names = [...new Set([...Object.keys(before), ...Object.keys(after)])].sort();
871
+ return names
872
+ .map((name) => {
873
+ const a = before[name];
874
+ const b = after[name];
875
+ const unit = a?.unit ?? b?.unit ?? "number";
876
+ const comparable = a && b && a.unit === b.unit;
877
+ const delta = comparable ? b.value - a.value : null;
878
+ const better = b?.better ?? a?.better ?? "neutral";
879
+ const tone =
880
+ better === "neutral" ? "metric-neutral" : toneForDelta(delta, better === "lower");
881
+ return comparisonBars({
882
+ title: metricName(name, b?.label ?? a?.label),
883
+ before: a?.value,
884
+ after: b?.value,
885
+ beforeText: metricValue(a?.value, unit),
886
+ afterText: metricValue(b?.value, unit),
887
+ deltaText: comparable ? signedMetricValue(delta, unit) : "not paired",
888
+ deltaTone: tone,
889
+ });
890
+ })
891
+ .join("");
892
+ };
893
+
894
+ const subagentTotals = (activities) => {
895
+ const subagents = (activities ?? []).filter((activity) => activity.agent?.role === "subagent");
896
+ return {
897
+ count: subagents.length,
898
+ toolCalls: subagents.reduce((sum, activity) => sum + (activity.toolCalls?.total ?? 0), 0),
899
+ tokens: subagents.reduce((sum, activity) => sum + (activity.usage?.total ?? 0), 0),
900
+ durationMs: subagents.reduce((sum, activity) => sum + (activity.durationMs ?? 0), 0),
901
+ };
902
+ };
903
+
904
+ export const renderRunComparison = (payload, labels) => {
905
+ const comparison = payload.comparison;
906
+ const runDelta = comparison.runDelta;
907
+ const durationDelta =
908
+ typeof runDelta?.durationMsA === "number" && typeof runDelta?.durationMsB === "number"
909
+ ? runDelta.durationMsB - runDelta.durationMsA
910
+ : null;
911
+ const mutationA = runDelta?.actionTiming?.a?.timeToFirstMutationMs;
912
+ const mutationB = runDelta?.actionTiming?.b?.timeToFirstMutationMs;
913
+ const mutationDelta =
914
+ typeof mutationA === "number" && typeof mutationB === "number" ? mutationB - mutationA : null;
915
+ const tokenA =
916
+ runDelta?.agentUsageAvailable?.a === false ? null : totalTokens(runDelta?.usage?.a?.agent);
917
+ const tokenB =
918
+ runDelta?.agentUsageAvailable?.b === false ? null : totalTokens(runDelta?.usage?.b?.agent);
919
+ const tokenDelta = tokenA === null || tokenB === null ? null : tokenB - tokenA;
920
+ const scoreDelta =
921
+ comparison.grading.caseScoreComparable !== false &&
922
+ typeof comparison.grading.caseScoreA === "number" &&
923
+ typeof comparison.grading.caseScoreB === "number"
924
+ ? comparison.grading.caseScoreB - comparison.grading.caseScoreA
925
+ : null;
926
+ let comparisonStatus = "ungraded";
927
+ if (payload.regressed === true) comparisonStatus = "failed";
928
+ if (payload.regressed === false) comparisonStatus = "passed";
929
+ const driverMetrics = driverMetricComparisons(runDelta);
930
+ const subagentsA = subagentTotals(runDelta?.agentActivity?.a);
931
+ const subagentsB = subagentTotals(runDelta?.agentActivity?.b);
932
+ return `
933
+ <header class="detail-header comparison-header">
934
+ <div class="detail-kicker">${badge(comparisonStatus)}<span>run comparison</span></div>
935
+ <div class="detail-title-row">
936
+ <div><h2>${escapeHtml(labels.before)} → ${escapeHtml(
937
+ labels.after,
938
+ )}</h2><p class="detail-subtitle">Baseline A compared with candidate B</p></div>
939
+ <button class="secondary-button" type="button" data-close-compare>Done</button>
940
+ </div>
941
+ </header>
942
+ <div class="detail-body">
943
+ ${
944
+ comparison.caseMismatch.length
945
+ ? `<div class="attention-banner">Evidence identity differs: ${escapeHtml(
946
+ comparison.caseMismatch.map((entry) => entry.field).join(", "),
947
+ )}. Treat deltas as potentially incomparable.</div>`
948
+ : '<div class="integrity-banner">✓ Case, evaluator, runtime, and driver identities are comparable</div>'
949
+ }
950
+ ${comparison.grading.caseScoreComparable === false ? '<div class="attention-banner">Score change unavailable: scored assertions or weights differ, or evidence is missing.</div>' : ""}
951
+ <dl class="comparison-stats">
952
+ <div><dt>Case score</dt><dd>${escapeHtml(
953
+ `${percentage(comparison.grading.caseScoreA)} → ${percentage(
954
+ comparison.grading.caseScoreB,
955
+ )}`,
956
+ )}</dd><small class="${toneForDelta(scoreDelta)}">${escapeHtml(
957
+ signedPercentage(scoreDelta),
958
+ )}</small></div>
959
+ <div><dt>Duration</dt><dd>${escapeHtml(
960
+ `${duration(runDelta?.durationMsA)} → ${duration(runDelta?.durationMsB)}`,
961
+ )}</dd><small class="${toneForDelta(durationDelta, true)}">${escapeHtml(
962
+ signedDuration(durationDelta),
963
+ )}</small></div>
964
+ <div><dt>First mutation</dt><dd>${escapeHtml(
965
+ `${duration(mutationA)} → ${duration(mutationB)}`,
966
+ )}</dd><small class="${toneForDelta(mutationDelta, true)}">${escapeHtml(
967
+ signedDuration(mutationDelta),
968
+ )}</small></div>
969
+ <div><dt>Agent tokens</dt><dd>${escapeHtml(
970
+ `${compactNumber(tokenA)} → ${compactNumber(tokenB)}`,
971
+ )}</dd><small class="${toneForDelta(tokenDelta, true)}">${tokenDelta > 0 ? "+" : ""}${escapeHtml(
972
+ compactNumber(tokenDelta),
973
+ )}</small></div>
974
+ <div><dt>Assertions</dt><dd>${comparison.grading.improvements} ↑ / ${comparison.grading.regressions} ↓</dd><small>${comparison.grading.deltas.length} compared</small></div>
975
+ </dl>
976
+ <section class="detail-section">
977
+ <h3>Comparison profile <span>shared scales within each measure</span></h3>
978
+ <div class="comparison-plot-grid">
979
+ ${comparisonBars({
980
+ title: "Case score",
981
+ continuous: true,
982
+ before: comparison.grading.caseScoreA,
983
+ after: comparison.grading.caseScoreB,
984
+ maximum: 1,
985
+ beforeText: percentage(comparison.grading.caseScoreA),
986
+ afterText: percentage(comparison.grading.caseScoreB),
987
+ deltaText: signedPercentage(scoreDelta),
988
+ deltaTone: toneForDelta(scoreDelta),
989
+ })}
990
+ ${comparisonBars({
991
+ title: "Duration",
992
+ before: runDelta?.durationMsA,
993
+ after: runDelta?.durationMsB,
994
+ beforeText: duration(runDelta?.durationMsA),
995
+ afterText: duration(runDelta?.durationMsB),
996
+ deltaText: signedDuration(durationDelta),
997
+ deltaTone: toneForDelta(durationDelta, true),
998
+ })}
999
+ ${comparisonBars({
1000
+ title: "Time to first mutation",
1001
+ before: mutationA,
1002
+ after: mutationB,
1003
+ beforeText: duration(mutationA),
1004
+ afterText: duration(mutationB),
1005
+ deltaText: signedDuration(mutationDelta),
1006
+ deltaTone: toneForDelta(mutationDelta, true),
1007
+ })}
1008
+ ${comparisonBars({
1009
+ title: "Agent tokens",
1010
+ before: tokenA,
1011
+ after: tokenB,
1012
+ beforeText: compactNumber(tokenA),
1013
+ afterText: compactNumber(tokenB),
1014
+ deltaText: `${tokenDelta > 0 ? "+" : ""}${compactNumber(tokenDelta)}`,
1015
+ deltaTone: toneForDelta(tokenDelta, true),
1016
+ })}
1017
+ ${comparisonBars({
1018
+ title: "Subagents spawned",
1019
+ before: subagentsA.count,
1020
+ after: subagentsB.count,
1021
+ beforeText: compactNumber(subagentsA.count),
1022
+ afterText: compactNumber(subagentsB.count),
1023
+ deltaText: `${subagentsB.count - subagentsA.count > 0 ? "+" : ""}${compactNumber(
1024
+ subagentsB.count - subagentsA.count,
1025
+ )}`,
1026
+ deltaTone: "metric-neutral",
1027
+ })}
1028
+ ${comparisonBars({
1029
+ title: "Subagent tool calls",
1030
+ before: subagentsA.toolCalls,
1031
+ after: subagentsB.toolCalls,
1032
+ beforeText: compactNumber(subagentsA.toolCalls),
1033
+ afterText: compactNumber(subagentsB.toolCalls),
1034
+ deltaText: `${subagentsB.toolCalls - subagentsA.toolCalls > 0 ? "+" : ""}${compactNumber(
1035
+ subagentsB.toolCalls - subagentsA.toolCalls,
1036
+ )}`,
1037
+ deltaTone: "metric-neutral",
1038
+ })}
1039
+ ${comparisonBars({
1040
+ title: "Total subagent time",
1041
+ before: subagentsA.durationMs,
1042
+ after: subagentsB.durationMs,
1043
+ beforeText: duration(subagentsA.durationMs),
1044
+ afterText: duration(subagentsB.durationMs),
1045
+ deltaText: signedDuration(subagentsB.durationMs - subagentsA.durationMs),
1046
+ deltaTone: "metric-neutral",
1047
+ })}
1048
+ ${comparisonBars({
1049
+ title: "Subagent tokens",
1050
+ before: subagentsA.tokens,
1051
+ after: subagentsB.tokens,
1052
+ beforeText: compactNumber(subagentsA.tokens),
1053
+ afterText: compactNumber(subagentsB.tokens),
1054
+ deltaText: `${subagentsB.tokens - subagentsA.tokens > 0 ? "+" : ""}${compactNumber(
1055
+ subagentsB.tokens - subagentsA.tokens,
1056
+ )}`,
1057
+ deltaTone: toneForDelta(subagentsB.tokens - subagentsA.tokens, true),
1058
+ })}
1059
+ </div>
1060
+ <div class="comparison-distributions">
1061
+ ${compositionFigure(
1062
+ "Assertion movement",
1063
+ [
1064
+ { key: "improved", label: "Improved", value: comparison.grading.improvements },
1065
+ { key: "regressed", label: "Regressed", value: comparison.grading.regressions },
1066
+ {
1067
+ key: "added",
1068
+ label: "Added",
1069
+ value: comparison.grading.deltas.filter((delta) => delta.status === "added").length,
1070
+ },
1071
+ {
1072
+ key: "removed",
1073
+ label: "Removed",
1074
+ value: comparison.grading.deltas.filter((delta) => delta.status === "removed")
1075
+ .length,
1076
+ },
1077
+ {
1078
+ key: "unavailable",
1079
+ label: "Unavailable",
1080
+ value: comparison.grading.deltas.filter((delta) => delta.status === "unavailable")
1081
+ .length,
1082
+ },
1083
+ {
1084
+ key: "unchanged",
1085
+ label: "Unchanged",
1086
+ value: comparison.grading.deltas.filter((delta) => delta.status === "unchanged")
1087
+ .length,
1088
+ },
1089
+ ],
1090
+ "Distribution of assertion changes",
1091
+ )}
1092
+ ${compositionFigure(
1093
+ "Artifact movement",
1094
+ [
1095
+ {
1096
+ key: "improved",
1097
+ label: "Added",
1098
+ value: comparison.artifactsDelta.added.length,
1099
+ },
1100
+ {
1101
+ key: "regressed",
1102
+ label: "Removed",
1103
+ value: comparison.artifactsDelta.removed.length,
1104
+ },
1105
+ {
1106
+ key: "changed",
1107
+ label: "Changed",
1108
+ value: comparison.artifactsDelta.changed.length,
1109
+ },
1110
+ {
1111
+ key: "unchanged",
1112
+ label: "Unchanged",
1113
+ value: comparison.artifactsDelta.unchanged.length,
1114
+ },
1115
+ ],
1116
+ "Distribution of artifact changes",
1117
+ )}
1118
+ </div>
1119
+ </section>
1120
+ ${
1121
+ driverMetrics
1122
+ ? `<section class="detail-section"><h3>Driver metric comparison <span>shared scale per metric</span></h3><div class="comparison-plot-grid">${driverMetrics}</div></section>`
1123
+ : ""
1124
+ }
1125
+ <section class="detail-section">
1126
+ <h3>Judgment delta <span>rationale side by side</span></h3>
1127
+ <div class="matrix-wrap"><table class="matrix comparison-matrix">
1128
+ <thead><tr><th>Assertion</th><th>A</th><th>B</th><th>Change</th><th>Judge rationale A → B</th></tr></thead>
1129
+ <tbody>${assertionDeltaRows(payload) || '<tr><td colspan="5">No assertions.</td></tr>'}</tbody>
1130
+ </table></div>
1131
+ </section>
1132
+ <section class="detail-section">
1133
+ <h3>Artifacts <span>${comparison.artifactsDelta.changed.length} changed · ${comparison.artifactsDelta.added.length} added · ${comparison.artifactsDelta.removed.length} removed</span></h3>
1134
+ <div class="artifact-list">${artifactDelta(payload) || '<div class="artifact"><span>No artifact changes</span><small>unchanged</small></div>'}</div>
1135
+ </section>
1136
+ </div>`;
1137
+ };
1138
+
1139
+ const metricDeltaSummary = (metrics) =>
1140
+ Object.entries(metrics)
1141
+ .map(
1142
+ ([name, summary]) =>
1143
+ `<span><b>${escapeHtml(name)}</b> ${escapeHtml(
1144
+ metricValue(summary.mean, inferredMetricUnit(name, summary.unit)),
1145
+ )}</span>`,
1146
+ )
1147
+ .join("");
1148
+
1149
+ export const renderSuiteComparison = (payload, labels) => {
1150
+ const comparison = payload.comparison;
1151
+ const movement = { improved: 0, regressed: 0, unchanged: 0, missing: 0 };
1152
+ for (const entry of comparison.cases) {
1153
+ if (!entry.before || !entry.after) movement.missing++;
1154
+ else if (entry.regressed) movement.regressed++;
1155
+ else if ((entry.passRateDelta ?? 0) > 0) movement.improved++;
1156
+ else movement.unchanged++;
1157
+ }
1158
+ return `
1159
+ <header class="detail-header comparison-header">
1160
+ <div class="detail-kicker">${badge(
1161
+ comparison.regressions || comparison.identityMismatches.length ? "failed" : "passed",
1162
+ )}<span>suite comparison</span></div>
1163
+ <div class="detail-title-row">
1164
+ <div><h2>${escapeHtml(labels.before)} → ${escapeHtml(
1165
+ labels.after,
1166
+ )}</h2><p class="detail-subtitle">Repeated evidence with paired case metrics</p></div>
1167
+ <button class="secondary-button" type="button" data-close-compare>Done</button>
1168
+ </div>
1169
+ </header>
1170
+ <div class="detail-body">
1171
+ ${
1172
+ comparison.identityMismatches.length
1173
+ ? `<div class="attention-banner">${comparison.identityMismatches.length} repetition identities differ; those pairs are not directly comparable.</div>`
1174
+ : '<div class="integrity-banner">✓ All paired repetition identities match</div>'
1175
+ }
1176
+ <dl class="comparison-stats">
1177
+ <div><dt>Pass rate</dt><dd>${escapeHtml(
1178
+ `${percentage(comparison.before.passRate)} → ${percentage(comparison.after.passRate)}`,
1179
+ )}</dd><small class="${toneForDelta(comparison.passRateDelta)}">${escapeHtml(
1180
+ signedPercentage(comparison.passRateDelta),
1181
+ )}</small></div>
1182
+ <div><dt>Attempts</dt><dd>${comparison.before.attempts} → ${comparison.after.attempts}</dd><small>paired populations</small></div>
1183
+ <div><dt>Regressions</dt><dd>${comparison.regressions}</dd><small>case-level</small></div>
1184
+ <div><dt>Identity drift</dt><dd>${comparison.identityMismatches.length}</dd><small>repetitions</small></div>
1185
+ </dl>
1186
+ <section class="detail-section">
1187
+ <h3>Outcome movement <span>shared 0–100% scale</span></h3>
1188
+ <div class="suite-outcome-visual">
1189
+ ${comparisonBars({
1190
+ title: "Suite pass rate",
1191
+ before: comparison.before.passRate,
1192
+ after: comparison.after.passRate,
1193
+ maximum: 1,
1194
+ beforeText: percentage(comparison.before.passRate),
1195
+ afterText: percentage(comparison.after.passRate),
1196
+ deltaText: signedPercentage(comparison.passRateDelta),
1197
+ deltaTone: toneForDelta(comparison.passRateDelta),
1198
+ })}
1199
+ ${compositionFigure(
1200
+ "Case movement",
1201
+ [
1202
+ {
1203
+ key: "improved",
1204
+ label: "Improved",
1205
+ value: movement.improved,
1206
+ },
1207
+ {
1208
+ key: "regressed",
1209
+ label: "Regressed",
1210
+ value: movement.regressed,
1211
+ },
1212
+ {
1213
+ key: "unchanged",
1214
+ label: "Unchanged",
1215
+ value: movement.unchanged,
1216
+ },
1217
+ {
1218
+ key: "pending",
1219
+ label: "Missing side",
1220
+ value: movement.missing,
1221
+ },
1222
+ ],
1223
+ "Distribution of per-case suite movement",
1224
+ )}
1225
+ </div>
1226
+ </section>
1227
+ <section class="detail-section">
1228
+ <h3>Per-case delta <span>after − before</span></h3>
1229
+ <div class="matrix-wrap"><table class="matrix comparison-matrix">
1230
+ <thead><tr><th>Case</th><th>Before</th><th>After</th><th>Δ</th><th>Paired metric means</th></tr></thead>
1231
+ <tbody>${comparison.cases
1232
+ .map(
1233
+ (entry) => `<tr class="${entry.regressed ? "row-regressed" : ""}">
1234
+ <td>${escapeHtml(entry.canonicalCaseId)}</td>
1235
+ <td class="bar-table-cell">${quantizedBar(
1236
+ entry.before?.passRate,
1237
+ "tone-before",
1238
+ `${entry.canonicalCaseId} baseline pass rate ${percentage(
1239
+ entry.before?.passRate,
1240
+ )}`,
1241
+ )}<small>${escapeHtml(percentage(entry.before?.passRate))} · n=${
1242
+ entry.before?.attempts ?? 0
1243
+ }</small></td>
1244
+ <td class="bar-table-cell">${quantizedBar(
1245
+ entry.after?.passRate,
1246
+ "tone-after",
1247
+ `${entry.canonicalCaseId} candidate pass rate ${percentage(
1248
+ entry.after?.passRate,
1249
+ )}`,
1250
+ )}<small>${escapeHtml(percentage(entry.after?.passRate))} · n=${
1251
+ entry.after?.attempts ?? 0
1252
+ }</small></td>
1253
+ <td class="${toneForDelta(entry.passRateDelta)}">${escapeHtml(
1254
+ signedPercentage(entry.passRateDelta),
1255
+ )}</td>
1256
+ <td class="metric-deltas">${metricDeltaSummary(entry.pairedMetricDeltas)}</td>
1257
+ </tr>`,
1258
+ )
1259
+ .join("")}</tbody>
1260
+ </table></div>
1261
+ </section>
1262
+ </div>`;
1263
+ };
1264
+
1265
+ export { renderArtifactComparison, renderArtifactPreview } from "./artifact-view.js";
1266
+ export { detailEmpty };