@tangle-network/agent-eval 0.129.0 → 0.130.0

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 (427) hide show
  1. package/CHANGELOG.md +14 -0
  2. package/README.md +1 -0
  3. package/dist/active-curriculum-C4mk67HP.js +214 -0
  4. package/dist/active-curriculum-C4mk67HP.js.map +1 -0
  5. package/dist/adversarial-smnADNFS.d.ts +21 -0
  6. package/dist/adversarial-smnADNFS.d.ts.map +1 -0
  7. package/dist/analyst/index.d.ts +81 -2872
  8. package/dist/analyst/index.d.ts.map +1 -0
  9. package/dist/analyst/index.js +319 -360
  10. package/dist/analyst/index.js.map +1 -1
  11. package/dist/analyst-BkTS3C58.d.ts +89 -0
  12. package/dist/analyst-BkTS3C58.d.ts.map +1 -0
  13. package/dist/analyst-LsnNpSkm.js +152 -0
  14. package/dist/analyst-LsnNpSkm.js.map +1 -0
  15. package/dist/analyze-runs-C1CavBMk.js +1067 -0
  16. package/dist/analyze-runs-C1CavBMk.js.map +1 -0
  17. package/dist/analyze-runs-FsgYCinh.d.ts +72 -0
  18. package/dist/analyze-runs-FsgYCinh.d.ts.map +1 -0
  19. package/dist/attribute-vocabulary-DLJ6303h.d.ts +54 -0
  20. package/dist/attribute-vocabulary-DLJ6303h.d.ts.map +1 -0
  21. package/dist/authenticity/index.d.ts +81 -79
  22. package/dist/authenticity/index.d.ts.map +1 -0
  23. package/dist/authenticity/index.js +209 -193
  24. package/dist/authenticity/index.js.map +1 -1
  25. package/dist/baseline-HsBvw_dk.js +550 -0
  26. package/dist/baseline-HsBvw_dk.js.map +1 -0
  27. package/dist/baseline-hG3K85h4.d.ts +125 -0
  28. package/dist/baseline-hG3K85h4.d.ts.map +1 -0
  29. package/dist/belief-state/index.d.ts +448 -1188
  30. package/dist/belief-state/index.d.ts.map +1 -0
  31. package/dist/belief-state/index.js +1617 -1709
  32. package/dist/belief-state/index.js.map +1 -1
  33. package/dist/benchmarks/index.d.ts +2 -891
  34. package/dist/benchmarks/index.js +2 -60
  35. package/dist/benchmarks-DviOvUNr.js +754 -0
  36. package/dist/benchmarks-DviOvUNr.js.map +1 -0
  37. package/dist/builder-eval/index.d.ts +150 -662
  38. package/dist/builder-eval/index.d.ts.map +1 -0
  39. package/dist/builder-eval/index.js +356 -345
  40. package/dist/builder-eval/index.js.map +1 -1
  41. package/dist/calibration-CNWWA6K8.js +94 -0
  42. package/dist/calibration-CNWWA6K8.js.map +1 -0
  43. package/dist/campaign/index.d.ts +5 -6381
  44. package/dist/campaign/index.js +3 -213
  45. package/dist/campaign-CBKZvQ1H.js +3885 -0
  46. package/dist/campaign-CBKZvQ1H.js.map +1 -0
  47. package/dist/cli.d.ts +1 -1
  48. package/dist/cli.js +164 -175
  49. package/dist/cli.js.map +1 -1
  50. package/dist/client-C97NMzqi.d.ts +581 -0
  51. package/dist/client-C97NMzqi.d.ts.map +1 -0
  52. package/dist/client-CYzbdJOZ.js +637 -0
  53. package/dist/client-CYzbdJOZ.js.map +1 -0
  54. package/dist/code-agent-session-BjkMTQ7H.js +1390 -0
  55. package/dist/code-agent-session-BjkMTQ7H.js.map +1 -0
  56. package/dist/code-agent-session-DqqgOJaz.d.ts +143 -0
  57. package/dist/code-agent-session-DqqgOJaz.d.ts.map +1 -0
  58. package/dist/concurrency-DIxRZF_J.js +85 -0
  59. package/dist/concurrency-DIxRZF_J.js.map +1 -0
  60. package/dist/contract/index.d.ts +645 -5565
  61. package/dist/contract/index.d.ts.map +1 -0
  62. package/dist/contract/index.js +1730 -1938
  63. package/dist/contract/index.js.map +1 -1
  64. package/dist/control.d.ts +3 -1030
  65. package/dist/control.js +2 -33
  66. package/dist/cost-ledger-DIgQUFZZ.js +801 -0
  67. package/dist/cost-ledger-DIgQUFZZ.js.map +1 -0
  68. package/dist/cost-ledger-Dye6jCgg.d.ts +236 -0
  69. package/dist/cost-ledger-Dye6jCgg.d.ts.map +1 -0
  70. package/dist/dataset-BvtnC8Dc.d.ts +115 -0
  71. package/dist/dataset-BvtnC8Dc.d.ts.map +1 -0
  72. package/dist/default-registry-C-vFCSEc.js +2579 -0
  73. package/dist/default-registry-C-vFCSEc.js.map +1 -0
  74. package/dist/default-registry-CNPo-Vsb.d.ts +540 -0
  75. package/dist/default-registry-CNPo-Vsb.d.ts.map +1 -0
  76. package/dist/emitter-CPBAhxum.js +266 -0
  77. package/dist/emitter-CPBAhxum.js.map +1 -0
  78. package/dist/emitter-DGQGoLyj.d.ts +113 -0
  79. package/dist/emitter-DGQGoLyj.d.ts.map +1 -0
  80. package/dist/errors-8YnH8WlF.js +64 -0
  81. package/dist/errors-8YnH8WlF.js.map +1 -0
  82. package/dist/errors-CEk209JS.d.ts +76 -0
  83. package/dist/errors-CEk209JS.d.ts.map +1 -0
  84. package/dist/eval-campaign-DEm6c8ru.js +349 -0
  85. package/dist/eval-campaign-DEm6c8ru.js.map +1 -0
  86. package/dist/execution-tracks-CpgFPpS5.js +93 -0
  87. package/dist/execution-tracks-CpgFPpS5.js.map +1 -0
  88. package/dist/exporters-q9iL-2Jf.js +148 -0
  89. package/dist/exporters-q9iL-2Jf.js.map +1 -0
  90. package/dist/extract-usage-BrQ8mCLX.js +155 -0
  91. package/dist/extract-usage-BrQ8mCLX.js.map +1 -0
  92. package/dist/failure-cluster-CqcvCcdR.d.ts +59 -0
  93. package/dist/failure-cluster-CqcvCcdR.d.ts.map +1 -0
  94. package/dist/feedback-trajectory-CVaeREXV.d.ts +340 -0
  95. package/dist/feedback-trajectory-CVaeREXV.d.ts.map +1 -0
  96. package/dist/fuzz.d.ts +320 -646
  97. package/dist/fuzz.d.ts.map +1 -0
  98. package/dist/fuzz.js +670 -618
  99. package/dist/fuzz.js.map +1 -1
  100. package/dist/hf-dataset-DBJXXoY1.js +763 -0
  101. package/dist/hf-dataset-DBJXXoY1.js.map +1 -0
  102. package/dist/hosted/index.d.ts +11 -831
  103. package/dist/hosted/index.d.ts.map +1 -0
  104. package/dist/hosted/index.js +2 -37
  105. package/dist/index-2JJSA6-r2.d.ts +926 -0
  106. package/dist/index-2JJSA6-r2.d.ts.map +1 -0
  107. package/dist/index-6N0aYmpW.d.ts +217 -0
  108. package/dist/index-6N0aYmpW.d.ts.map +1 -0
  109. package/dist/index-BAvgST_9.d.ts +131 -0
  110. package/dist/index-BAvgST_9.d.ts.map +1 -0
  111. package/dist/index-BvnJuTGD.d.ts +68 -0
  112. package/dist/index-BvnJuTGD.d.ts.map +1 -0
  113. package/dist/index-C61Wi7yg.d.ts +547 -0
  114. package/dist/index-C61Wi7yg.d.ts.map +1 -0
  115. package/dist/index-CAPUUKaM.d.ts +335 -0
  116. package/dist/index-CAPUUKaM.d.ts.map +1 -0
  117. package/dist/index-DE5fb3EC.d.ts +2244 -0
  118. package/dist/index-DE5fb3EC.d.ts.map +1 -0
  119. package/dist/index-DSC51roc.d.ts +102 -0
  120. package/dist/index-DSC51roc.d.ts.map +1 -0
  121. package/dist/index.d.ts +3755 -15555
  122. package/dist/index.d.ts.map +1 -0
  123. package/dist/index.js +11182 -11216
  124. package/dist/index.js.map +1 -1
  125. package/dist/integrity-BzRbCHzi.js +424 -0
  126. package/dist/integrity-BzRbCHzi.js.map +1 -0
  127. package/dist/integrity-rmVhXWA7.d.ts +61 -0
  128. package/dist/integrity-rmVhXWA7.d.ts.map +1 -0
  129. package/dist/judge-calibration-DFtEMlde.d.ts +146 -0
  130. package/dist/judge-calibration-DFtEMlde.d.ts.map +1 -0
  131. package/dist/ledger-core/index.d.ts +2 -0
  132. package/dist/ledger-core/index.js +2 -0
  133. package/dist/ledger-core-DtZz1RG0.js +388 -0
  134. package/dist/ledger-core-DtZz1RG0.js.map +1 -0
  135. package/dist/llm-client--GR4JbZE.js +687 -0
  136. package/dist/llm-client--GR4JbZE.js.map +1 -0
  137. package/dist/llm-client-B_nIBlYo.d.ts +290 -0
  138. package/dist/llm-client-B_nIBlYo.d.ts.map +1 -0
  139. package/dist/matrix/index.d.ts +3 -155
  140. package/dist/matrix/index.js +2 -8
  141. package/dist/matrix-BzQnu2S6.js +270 -0
  142. package/dist/matrix-BzQnu2S6.js.map +1 -0
  143. package/dist/meta-eval/index.d.ts +4 -1027
  144. package/dist/meta-eval/index.js +393 -390
  145. package/dist/meta-eval/index.js.map +1 -1
  146. package/dist/metrics-C9YY1OcL.js +239 -0
  147. package/dist/metrics-C9YY1OcL.js.map +1 -0
  148. package/dist/mint-yN2M2eh0.js +201 -0
  149. package/dist/mint-yN2M2eh0.js.map +1 -0
  150. package/dist/multi-layer-verifier-BHY1gWAc.d.ts +138 -0
  151. package/dist/multi-layer-verifier-BHY1gWAc.d.ts.map +1 -0
  152. package/dist/multishot/index.d.ts +273 -480
  153. package/dist/multishot/index.d.ts.map +1 -0
  154. package/dist/multishot/index.js +595 -548
  155. package/dist/multishot/index.js.map +1 -1
  156. package/dist/off-policy-DvgzvtIx.js +220 -0
  157. package/dist/off-policy-DvgzvtIx.js.map +1 -0
  158. package/dist/off-policy-mskQw8Mb.d.ts +153 -0
  159. package/dist/off-policy-mskQw8Mb.d.ts.map +1 -0
  160. package/dist/openapi.json +1 -1
  161. package/dist/opencode-sqlite-BGrHeDu3.js +318 -0
  162. package/dist/opencode-sqlite-BGrHeDu3.js.map +1 -0
  163. package/dist/outcome-store-BYHIuO0e.d.ts +65 -0
  164. package/dist/outcome-store-BYHIuO0e.d.ts.map +1 -0
  165. package/dist/outcome-store-ChBKlTd_.js +75 -0
  166. package/dist/outcome-store-ChBKlTd_.js.map +1 -0
  167. package/dist/paired-arms-D9D0wXj2.js +260 -0
  168. package/dist/paired-arms-D9D0wXj2.js.map +1 -0
  169. package/dist/pipelines/index.d.ts +95 -532
  170. package/dist/pipelines/index.d.ts.map +1 -0
  171. package/dist/pipelines/index.js +497 -478
  172. package/dist/pipelines/index.js.map +1 -1
  173. package/dist/pre-registration-DakwTRXk.js +96 -0
  174. package/dist/pre-registration-DakwTRXk.js.map +1 -0
  175. package/dist/propose-review-control-LhIWGzHl.js +1458 -0
  176. package/dist/propose-review-control-LhIWGzHl.js.map +1 -0
  177. package/dist/query-CJ_DX8vl.d.ts +28 -0
  178. package/dist/query-CJ_DX8vl.d.ts.map +1 -0
  179. package/dist/query-Di7eEQ79.js +83 -0
  180. package/dist/query-Di7eEQ79.js.map +1 -0
  181. package/dist/raw-provider-sink-BQd7mzyT.js +201 -0
  182. package/dist/raw-provider-sink-BQd7mzyT.js.map +1 -0
  183. package/dist/raw-provider-sink-BU29Sh8h.d.ts +134 -0
  184. package/dist/raw-provider-sink-BU29Sh8h.d.ts.map +1 -0
  185. package/dist/redact-7Aq1ukl-.js +107 -0
  186. package/dist/redact-7Aq1ukl-.js.map +1 -0
  187. package/dist/release-report-Cz9NKH39.js +603 -0
  188. package/dist/release-report-Cz9NKH39.js.map +1 -0
  189. package/dist/release-report-mvB2G4_J.d.ts +244 -0
  190. package/dist/release-report-mvB2G4_J.d.ts.map +1 -0
  191. package/dist/replay-GnyotH0J.js +1741 -0
  192. package/dist/replay-GnyotH0J.js.map +1 -0
  193. package/dist/replay-OoidtG1E.d.ts +749 -0
  194. package/dist/replay-OoidtG1E.d.ts.map +1 -0
  195. package/dist/reporting.d.ts +6 -1312
  196. package/dist/reporting.js +6 -51
  197. package/dist/researcher-CwTdwXG1.d.ts +314 -0
  198. package/dist/researcher-CwTdwXG1.d.ts.map +1 -0
  199. package/dist/reward-hacking-eAnOsynk.d.ts +224 -0
  200. package/dist/reward-hacking-eAnOsynk.d.ts.map +1 -0
  201. package/dist/reward-hacking-qipEpKvY.js +596 -0
  202. package/dist/reward-hacking-qipEpKvY.js.map +1 -0
  203. package/dist/reward-nw2xZGZG.js +137 -0
  204. package/dist/reward-nw2xZGZG.js.map +1 -0
  205. package/dist/rl.d.ts +760 -4010
  206. package/dist/rl.d.ts.map +1 -0
  207. package/dist/rl.js +2325 -1958
  208. package/dist/rl.js.map +1 -1
  209. package/dist/rolldown-runtime-8H4AJuhK.js +14 -0
  210. package/dist/rollout/index.d.ts +3 -2087
  211. package/dist/rollout/index.js +8 -168
  212. package/dist/rollout-BOYjemfR.js +624 -0
  213. package/dist/rollout-BOYjemfR.js.map +1 -0
  214. package/dist/rubric-predictive-validity-B3xmbmS1.js +141 -0
  215. package/dist/rubric-predictive-validity-B3xmbmS1.js.map +1 -0
  216. package/dist/rubric-predictive-validity-Ku_clp_1.d.ts +74 -0
  217. package/dist/rubric-predictive-validity-Ku_clp_1.d.ts.map +1 -0
  218. package/dist/run-evidence-ClRX_8A9.d.ts +225 -0
  219. package/dist/run-evidence-ClRX_8A9.d.ts.map +1 -0
  220. package/dist/run-record-BuoE80Dq.js +467 -0
  221. package/dist/{chunk-56TAVBOK.js.map → run-record-BuoE80Dq.js.map} +1 -1
  222. package/dist/run-record-CnZu_gjl.d.ts +357 -0
  223. package/dist/run-record-CnZu_gjl.d.ts.map +1 -0
  224. package/dist/run-score-iEEAWiBY.js +41 -0
  225. package/dist/run-score-iEEAWiBY.js.map +1 -0
  226. package/dist/runtime-trajectory-1gyaTOoC.js +93 -0
  227. package/dist/runtime-trajectory-1gyaTOoC.js.map +1 -0
  228. package/dist/runtime-trajectory-BvSZcCHD.d.ts +50 -0
  229. package/dist/runtime-trajectory-BvSZcCHD.d.ts.map +1 -0
  230. package/dist/schema-BtVldJ3T.d.ts +206 -0
  231. package/dist/schema-BtVldJ3T.d.ts.map +1 -0
  232. package/dist/schema-C6DW4ZHR.js +821 -0
  233. package/dist/schema-C6DW4ZHR.js.map +1 -0
  234. package/dist/schema-CRhEY1SO.js +69 -0
  235. package/dist/schema-CRhEY1SO.js.map +1 -0
  236. package/dist/schema-Cef2cFmb.d.ts +408 -0
  237. package/dist/schema-Cef2cFmb.d.ts.map +1 -0
  238. package/dist/semantic-concept-judge-B6cWNJ2K.js +725 -0
  239. package/dist/semantic-concept-judge-B6cWNJ2K.js.map +1 -0
  240. package/dist/sequential-Br0mAPHA.js +148 -0
  241. package/dist/sequential-Br0mAPHA.js.map +1 -0
  242. package/dist/sequential-CYwq6Ff_.d.ts +141 -0
  243. package/dist/sequential-CYwq6Ff_.d.ts.map +1 -0
  244. package/dist/series-convergence-CjO2QdRW.js +43 -0
  245. package/dist/series-convergence-CjO2QdRW.js.map +1 -0
  246. package/dist/series-convergence-ofsqPWhs.d.ts +35 -0
  247. package/dist/series-convergence-ofsqPWhs.d.ts.map +1 -0
  248. package/dist/server-m5D9cvnG.js +1040 -0
  249. package/dist/server-m5D9cvnG.js.map +1 -0
  250. package/dist/skill-usage-C_pXm7lP.d.ts +534 -0
  251. package/dist/skill-usage-C_pXm7lP.d.ts.map +1 -0
  252. package/dist/skillopt-optimization-method-B7wX7XkF.d.ts +1739 -0
  253. package/dist/skillopt-optimization-method-B7wX7XkF.d.ts.map +1 -0
  254. package/dist/skillopt-optimization-method-D4ODwFVV.js +7712 -0
  255. package/dist/skillopt-optimization-method-D4ODwFVV.js.map +1 -0
  256. package/dist/statistics-Cmj6nynr.d.ts +514 -0
  257. package/dist/statistics-Cmj6nynr.d.ts.map +1 -0
  258. package/dist/statistics-CnnxdpOg.js +1437 -0
  259. package/dist/statistics-CnnxdpOg.js.map +1 -0
  260. package/dist/store-CT9YIIve.d.ts +117 -0
  261. package/dist/store-CT9YIIve.d.ts.map +1 -0
  262. package/dist/store-CxJry_cs.d.ts +229 -0
  263. package/dist/store-CxJry_cs.d.ts.map +1 -0
  264. package/dist/storyboard/index.d.ts +84 -203
  265. package/dist/storyboard/index.d.ts.map +1 -0
  266. package/dist/storyboard/index.js +609 -542
  267. package/dist/storyboard/index.js.map +1 -1
  268. package/dist/summary-report-BNs5nmXI.js +862 -0
  269. package/dist/summary-report-BNs5nmXI.js.map +1 -0
  270. package/dist/summary-report-Cj9gdw4i.d.ts +407 -0
  271. package/dist/summary-report-Cj9gdw4i.d.ts.map +1 -0
  272. package/dist/supervisor-run/index.d.ts +2 -959
  273. package/dist/supervisor-run/index.js +2 -65
  274. package/dist/supervisor-run-_lnTLM3z.js +1679 -0
  275. package/dist/supervisor-run-_lnTLM3z.js.map +1 -0
  276. package/dist/task-failure-attributes-CQZlB3et.js +311 -0
  277. package/dist/task-failure-attributes-CQZlB3et.js.map +1 -0
  278. package/dist/test-graded-scenario-BsqWLmPt.js +318 -0
  279. package/dist/test-graded-scenario-BsqWLmPt.js.map +1 -0
  280. package/dist/test-graded-scenario-D1TaI2va.d.ts +141 -0
  281. package/dist/test-graded-scenario-D1TaI2va.d.ts.map +1 -0
  282. package/dist/tools-BmuN627J.js +1085 -0
  283. package/dist/tools-BmuN627J.js.map +1 -0
  284. package/dist/trace-attributes.d.ts +2 -52
  285. package/dist/trace-attributes.js +131 -61
  286. package/dist/trace-attributes.js.map +1 -1
  287. package/dist/traces.d.ts +12 -2365
  288. package/dist/traces.js +12 -252
  289. package/dist/trajectory-D_7rLrvE.js +56 -0
  290. package/dist/trajectory-D_7rLrvE.js.map +1 -0
  291. package/dist/types-DGsxbAEd.d.ts +387 -0
  292. package/dist/types-DGsxbAEd.d.ts.map +1 -0
  293. package/dist/types-k9tZGKUg.d.ts +640 -0
  294. package/dist/types-k9tZGKUg.d.ts.map +1 -0
  295. package/dist/verdict-Dps8_okt.d.ts +37 -0
  296. package/dist/verdict-Dps8_okt.d.ts.map +1 -0
  297. package/dist/wire/index.d.ts +702 -1173
  298. package/dist/wire/index.d.ts.map +1 -0
  299. package/dist/wire/index.js +2 -81
  300. package/package.json +17 -9
  301. package/dist/benchmarks/index.js.map +0 -1
  302. package/dist/campaign/index.js.map +0 -1
  303. package/dist/chunk-2QU3YOPR.js +0 -7374
  304. package/dist/chunk-2QU3YOPR.js.map +0 -1
  305. package/dist/chunk-3OCR4R5I.js +0 -728
  306. package/dist/chunk-3OCR4R5I.js.map +0 -1
  307. package/dist/chunk-3RF76KTD.js +0 -84
  308. package/dist/chunk-3RF76KTD.js.map +0 -1
  309. package/dist/chunk-56TAVBOK.js +0 -698
  310. package/dist/chunk-5DTSBUL2.js +0 -159
  311. package/dist/chunk-5DTSBUL2.js.map +0 -1
  312. package/dist/chunk-7FO3TNPI.js +0 -232
  313. package/dist/chunk-7FO3TNPI.js.map +0 -1
  314. package/dist/chunk-7ZZMD7UK.js +0 -386
  315. package/dist/chunk-7ZZMD7UK.js.map +0 -1
  316. package/dist/chunk-BOD4O7OF.js +0 -40
  317. package/dist/chunk-BOD4O7OF.js.map +0 -1
  318. package/dist/chunk-BSO5JDQH.js +0 -2335
  319. package/dist/chunk-BSO5JDQH.js.map +0 -1
  320. package/dist/chunk-C6LXANRU.js +0 -1550
  321. package/dist/chunk-C6LXANRU.js.map +0 -1
  322. package/dist/chunk-DODXQREJ.js +0 -752
  323. package/dist/chunk-DODXQREJ.js.map +0 -1
  324. package/dist/chunk-DRYIUNWY.js +0 -622
  325. package/dist/chunk-DRYIUNWY.js.map +0 -1
  326. package/dist/chunk-E7QXT7SX.js +0 -183
  327. package/dist/chunk-E7QXT7SX.js.map +0 -1
  328. package/dist/chunk-EG66UGL4.js +0 -341
  329. package/dist/chunk-EG66UGL4.js.map +0 -1
  330. package/dist/chunk-FXTVJPYD.js +0 -576
  331. package/dist/chunk-FXTVJPYD.js.map +0 -1
  332. package/dist/chunk-G7MGMCZD.js +0 -153
  333. package/dist/chunk-G7MGMCZD.js.map +0 -1
  334. package/dist/chunk-GGE4NNQT.js +0 -65
  335. package/dist/chunk-GGE4NNQT.js.map +0 -1
  336. package/dist/chunk-H23X7XKK.js +0 -181
  337. package/dist/chunk-H23X7XKK.js.map +0 -1
  338. package/dist/chunk-HHWE3POT.js +0 -94
  339. package/dist/chunk-HHWE3POT.js.map +0 -1
  340. package/dist/chunk-HPWUNB47.js +0 -289
  341. package/dist/chunk-HPWUNB47.js.map +0 -1
  342. package/dist/chunk-IYCLP2N2.js +0 -766
  343. package/dist/chunk-IYCLP2N2.js.map +0 -1
  344. package/dist/chunk-JHCHEVET.js +0 -274
  345. package/dist/chunk-JHCHEVET.js.map +0 -1
  346. package/dist/chunk-JQSF5DQT.js +0 -701
  347. package/dist/chunk-JQSF5DQT.js.map +0 -1
  348. package/dist/chunk-K4DBDHLK.js +0 -158
  349. package/dist/chunk-K4DBDHLK.js.map +0 -1
  350. package/dist/chunk-K6N6XJJX.js +0 -306
  351. package/dist/chunk-K6N6XJJX.js.map +0 -1
  352. package/dist/chunk-M4YBQKIJ.js +0 -1040
  353. package/dist/chunk-M4YBQKIJ.js.map +0 -1
  354. package/dist/chunk-MA6HLL3S.js +0 -65
  355. package/dist/chunk-MA6HLL3S.js.map +0 -1
  356. package/dist/chunk-MAZ26DC7.js +0 -99
  357. package/dist/chunk-MAZ26DC7.js.map +0 -1
  358. package/dist/chunk-NPCTHQIO.js +0 -91
  359. package/dist/chunk-NPCTHQIO.js.map +0 -1
  360. package/dist/chunk-NY44NC4A.js +0 -1056
  361. package/dist/chunk-NY44NC4A.js.map +0 -1
  362. package/dist/chunk-OIUOT4QD.js +0 -44
  363. package/dist/chunk-OIUOT4QD.js.map +0 -1
  364. package/dist/chunk-ONWEPEDO.js +0 -57
  365. package/dist/chunk-ONWEPEDO.js.map +0 -1
  366. package/dist/chunk-OWN5NPMC.js +0 -152
  367. package/dist/chunk-OWN5NPMC.js.map +0 -1
  368. package/dist/chunk-P6FYH6K4.js +0 -1161
  369. package/dist/chunk-P6FYH6K4.js.map +0 -1
  370. package/dist/chunk-PC4UYEBM.js +0 -166
  371. package/dist/chunk-PC4UYEBM.js.map +0 -1
  372. package/dist/chunk-PC5DOSM7.js +0 -579
  373. package/dist/chunk-PC5DOSM7.js.map +0 -1
  374. package/dist/chunk-PXE2VKMX.js +0 -140
  375. package/dist/chunk-PXE2VKMX.js.map +0 -1
  376. package/dist/chunk-PZ5AY32C.js +0 -10
  377. package/dist/chunk-PZ5AY32C.js.map +0 -1
  378. package/dist/chunk-QB6BDBP2.js +0 -4464
  379. package/dist/chunk-QB6BDBP2.js.map +0 -1
  380. package/dist/chunk-RXHCETDZ.js +0 -536
  381. package/dist/chunk-RXHCETDZ.js.map +0 -1
  382. package/dist/chunk-RZTMDUO7.js +0 -49
  383. package/dist/chunk-RZTMDUO7.js.map +0 -1
  384. package/dist/chunk-SFLLL76A.js +0 -669
  385. package/dist/chunk-SFLLL76A.js.map +0 -1
  386. package/dist/chunk-SZLVEKMJ.js +0 -1446
  387. package/dist/chunk-SZLVEKMJ.js.map +0 -1
  388. package/dist/chunk-T4SQEITX.js +0 -95
  389. package/dist/chunk-T4SQEITX.js.map +0 -1
  390. package/dist/chunk-T6RLYGAD.js +0 -158
  391. package/dist/chunk-T6RLYGAD.js.map +0 -1
  392. package/dist/chunk-TJVT4QFF.js +0 -911
  393. package/dist/chunk-TJVT4QFF.js.map +0 -1
  394. package/dist/chunk-TQ7LNKZ3.js +0 -136
  395. package/dist/chunk-TQ7LNKZ3.js.map +0 -1
  396. package/dist/chunk-U4L7JRPZ.js +0 -1706
  397. package/dist/chunk-U4L7JRPZ.js.map +0 -1
  398. package/dist/chunk-U4PHLT2N.js +0 -419
  399. package/dist/chunk-U4PHLT2N.js.map +0 -1
  400. package/dist/chunk-VCZ5FQYW.js +0 -928
  401. package/dist/chunk-VCZ5FQYW.js.map +0 -1
  402. package/dist/chunk-VI2UW6B6.js +0 -162
  403. package/dist/chunk-VI2UW6B6.js.map +0 -1
  404. package/dist/chunk-VQMK5FMP.js +0 -247
  405. package/dist/chunk-VQMK5FMP.js.map +0 -1
  406. package/dist/chunk-WGXIEX7P.js +0 -116
  407. package/dist/chunk-WGXIEX7P.js.map +0 -1
  408. package/dist/chunk-WVATSFCP.js +0 -1553
  409. package/dist/chunk-WVATSFCP.js.map +0 -1
  410. package/dist/chunk-X4YIBDER.js +0 -1662
  411. package/dist/chunk-X4YIBDER.js.map +0 -1
  412. package/dist/chunk-YQN4ICPP.js +0 -355
  413. package/dist/chunk-YQN4ICPP.js.map +0 -1
  414. package/dist/chunk-ZET2UAYW.js +0 -89
  415. package/dist/chunk-ZET2UAYW.js.map +0 -1
  416. package/dist/chunk-ZHTZ4EYI.js +0 -1212
  417. package/dist/chunk-ZHTZ4EYI.js.map +0 -1
  418. package/dist/control.js.map +0 -1
  419. package/dist/hosted/index.js.map +0 -1
  420. package/dist/matrix/index.js.map +0 -1
  421. package/dist/reporting.js.map +0 -1
  422. package/dist/rollout/index.js.map +0 -1
  423. package/dist/run-campaign-OJJ7CZF4.js +0 -18
  424. package/dist/run-campaign-OJJ7CZF4.js.map +0 -1
  425. package/dist/supervisor-run/index.js.map +0 -1
  426. package/dist/traces.js.map +0 -1
  427. package/dist/wire/index.js.map +0 -1
@@ -1,121 +1,20 @@
1
- type SpanKind = 'agent' | 'llm' | 'tool' | 'retrieval' | 'judge' | 'sandbox' | 'custom';
2
- type SpanStatus = 'ok' | 'error';
3
- interface SpanBase {
4
- spanId: string;
5
- parentSpanId?: string;
6
- runId: string;
7
- kind: SpanKind;
8
- name: string;
9
- startedAt: number;
10
- endedAt?: number;
11
- status?: SpanStatus;
12
- error?: string;
13
- /** Anything not covered by typed fields. Kept deliberately free-form. */
14
- attributes?: Record<string, unknown>;
15
- }
16
- interface Message {
17
- role: 'system' | 'user' | 'assistant' | 'tool';
18
- content: string;
19
- tokens?: number;
20
- /** Multi-modal content descriptors; blobs themselves live in Artifacts. */
21
- images?: Array<{
22
- artifactId?: string;
23
- url?: string;
24
- mime?: string;
25
- }>;
26
- }
27
- interface LlmSpan extends SpanBase {
28
- kind: 'llm';
29
- model: string;
30
- messages: Message[];
31
- output?: string;
32
- inputTokens?: number;
33
- /** All generated tokens, including the reasoning subset when present. */
34
- outputTokens?: number;
35
- cachedTokens?: number;
36
- cacheWriteTokens?: number;
37
- /** Reasoning-token subset of `outputTokens`. */
38
- reasoningTokens?: number;
39
- costUsd?: number;
40
- finishReason?: string;
41
- }
42
- interface ToolSpan extends SpanBase {
43
- kind: 'tool';
44
- toolName: string;
45
- args: unknown;
46
- /** False when the source observed the call but did not capture its arguments. */
47
- argsCaptured?: boolean;
48
- result?: unknown;
49
- latencyMs?: number;
50
- }
51
- interface RetrievalSpan extends SpanBase {
52
- kind: 'retrieval';
53
- query: string;
54
- hits: Array<{
55
- docId: string;
56
- score: number;
57
- content?: string;
58
- }>;
59
- }
60
- interface JudgeSpan extends SpanBase {
61
- kind: 'judge';
62
- judgeId: string;
63
- /** Span this judgment applies to. */
64
- targetSpanId: string;
65
- dimension: string;
66
- /** Numeric score (free-range; interpretation up to the judge). */
67
- score: number;
68
- rationale?: string;
69
- evidence?: string;
70
- }
71
- interface SandboxSpan extends SpanBase {
72
- kind: 'sandbox';
73
- image?: string;
74
- command?: string;
75
- exitCode?: number;
76
- testsTotal?: number;
77
- testsPassed?: number;
78
- stdoutHash?: string;
79
- stderrHash?: string;
80
- /** Duration in ms; the harness fills this explicitly (endedAt - startedAt may miss setup). */
81
- wallMs?: number;
82
- }
83
- interface GenericSpan extends SpanBase {
84
- kind: 'agent' | 'custom';
85
- }
86
- type Span = LlmSpan | ToolSpan | RetrievalSpan | JudgeSpan | SandboxSpan | GenericSpan;
87
-
88
- /**
89
- * Code-edit extraction + the "watch the agent write code" animation.
90
- *
91
- * The base storyboard (./index) compresses a run into scenes but keeps scenes
92
- * payload-free (title + narration + evidence span ids). To actually SHOW the
93
- * code an agent wrote, this module pulls the concrete edit (path, diff,
94
- * before/after, language) out of the raw edit-tool spans and renders a
95
- * self-contained HTML replay that types each file out, character by character,
96
- * like an editor recording.
97
- *
98
- * Additive on purpose: it reuses the canonical `Span` (trace/schema) and the
99
- * `Storyboard` IR (./index) without modifying the reducer/compiler — so it
100
- * composes with the existing pipeline and a Remotion consumer can read the same
101
- * `CodeEdit[]` to render an MP4.
102
- */
103
-
1
+ import { _ as Span } from "../schema-BtVldJ3T.js";
2
+ //#region src/storyboard/code-edit.d.ts
104
3
  /** A concrete code change lifted from a tool span. */
105
4
  interface CodeEdit {
106
- /** Span the edit came from — back-reference into the trace. */
107
- spanId: string;
108
- path: string;
109
- language?: string;
110
- /** Unified diff, when the tool carried one. */
111
- diff?: string;
112
- /** Full file body before / after, when present — enables true keystroke
113
- * animation rather than only flashing a diff. */
114
- before?: string;
115
- after?: string;
116
- additions: number;
117
- deletions: number;
118
- startedAt: number;
5
+ /** Span the edit came from — back-reference into the trace. */
6
+ spanId: string;
7
+ path: string;
8
+ language?: string;
9
+ /** Unified diff, when the tool carried one. */
10
+ diff?: string;
11
+ /** Full file body before / after, when present — enables true keystroke
12
+ * animation rather than only flashing a diff. */
13
+ before?: string;
14
+ after?: string;
15
+ additions: number;
16
+ deletions: number;
17
+ startedAt: number;
119
18
  }
120
19
  /** Extract a CodeEdit from a single span, or undefined if it is not an edit. */
121
20
  declare function codeEditFromSpan(span: Span): CodeEdit | undefined;
@@ -124,20 +23,20 @@ declare function extractCodeEdits(spans: readonly Span[]): CodeEdit[];
124
23
  /** Pair each code-edit (`diff`) scene with its concrete edit, matched via the
125
24
  * scene's evidence span ids. Scenes without a recoverable edit are dropped. */
126
25
  declare function codeEditsForStoryboard(storyboard: Storyboard, spans: readonly Span[]): Array<{
127
- scene: Scene;
128
- edit: CodeEdit;
26
+ scene: Scene;
27
+ edit: CodeEdit;
129
28
  }>;
130
29
  /** The text the animation types for an edit: prefer the full new file body,
131
30
  * then the added lines of a diff, then a minimal placeholder so the scene is
132
31
  * never blank. */
133
32
  declare function editAnimationText(edit: CodeEdit): string;
134
33
  interface CodeAnimationOptions {
135
- /** Page title. Default "Agent writing code". */
136
- title?: string;
137
- /** Characters typed per animation tick. Higher = faster. Default 4. */
138
- charsPerTick?: number;
139
- /** Pause between files, ms. Default 900. */
140
- pauseBetweenMs?: number;
34
+ /** Page title. Default "Agent writing code". */
35
+ title?: string;
36
+ /** Characters typed per animation tick. Higher = faster. Default 4. */
37
+ charsPerTick?: number;
38
+ /** Pause between files, ms. Default 900. */
39
+ pauseBetweenMs?: number;
141
40
  }
142
41
  /**
143
42
  * Render a self-contained HTML page that animates the agent writing each file,
@@ -147,83 +46,64 @@ interface CodeAnimationOptions {
147
46
  * package; this is the dependency-free baseline.
148
47
  */
149
48
  declare function renderCodeAnimationHtml(source: readonly Span[] | readonly CodeEdit[], opts?: CodeAnimationOptions): string;
150
-
151
- /**
152
- * Storyboard — compile a raw agent trace into a small set of meaningful scenes,
153
- * then render them as a shareable, watchable replay.
154
- *
155
- * Span[] → SemanticEvent[] → Storyboard{scenes} → { markdown | html }
156
- *
157
- * The trace is the source of truth; the storyboard is the IR; markdown/html
158
- * (and, in a consumer package, Remotion/MP4) are compiled targets. Everything
159
- * here is PURE — same spans in, same bytes out (no clock/random) — so a run's
160
- * replay is deterministic and diffable.
161
- *
162
- * Crucially, every scene carries a modality-typed VISUAL extracted from the
163
- * span — a browser/computer screenshot, a code diff, a terminal command +
164
- * output, file content, an API request/response, or reasoning prose — so no
165
- * matter what the agent did, the replay SHOWS it, not just narrates it. The
166
- * reducer is rules-based; an LLM pass can refine titles/summaries later, but
167
- * the structure + visuals do not depend on one. Renderers that need a browser
168
- * (Remotion, React) live in consumers; this module only emits strings.
169
- */
170
-
49
+ //#endregion
50
+ //#region src/storyboard/index.d.ts
171
51
  /** What the agent meaningfully did — the compressed vocabulary. */
172
52
  type SemanticKind = 'understood_task' | 'user_message' | 'agent_reply' | 'reasoned' | 'ran_command' | 'read_file' | 'edited_code' | 'searched' | 'used_browser' | 'used_computer' | 'called_api' | 'called_tool' | 'evaluated' | 'observed_failure' | 'completed';
173
53
  /** The modality-typed payload a scene shows. This is what makes any agent
174
54
  * action — screen, browser, shell, code, API — actually viewable. */
175
55
  type SceneVisual = {
176
- type: 'screenshot';
177
- src: string;
178
- caption?: string;
56
+ type: 'screenshot';
57
+ src: string;
58
+ caption?: string;
179
59
  } | {
180
- type: 'browser';
181
- url?: string;
182
- action?: string;
183
- screenshot?: string;
60
+ type: 'browser';
61
+ url?: string;
62
+ action?: string;
63
+ screenshot?: string;
184
64
  } | {
185
- type: 'diff';
186
- path: string;
187
- patch: string;
65
+ type: 'diff';
66
+ path: string;
67
+ patch: string;
188
68
  } | {
189
- type: 'code';
190
- path: string;
191
- content: string;
69
+ type: 'code';
70
+ path: string;
71
+ content: string;
192
72
  } | {
193
- type: 'terminal';
194
- command: string;
195
- output?: string;
73
+ type: 'terminal';
74
+ command: string;
75
+ output?: string;
196
76
  } | {
197
- type: 'api';
198
- method?: string;
199
- url: string;
200
- status?: number;
201
- request?: string;
202
- response?: string;
77
+ type: 'api';
78
+ method?: string;
79
+ url: string;
80
+ status?: number;
81
+ request?: string;
82
+ response?: string;
203
83
  } | {
204
- type: 'prose';
205
- text: string;
84
+ type: 'prose';
85
+ text: string;
206
86
  } | {
207
- type: 'none';
87
+ type: 'none';
208
88
  };
209
89
  interface SemanticEvent {
210
- kind: SemanticKind;
211
- /** One-line headline (the ticket/scene title). */
212
- title: string;
213
- /** Short human summary. */
214
- summary: string;
215
- /** The modality payload to show for this moment. */
216
- visual: SceneVisual;
217
- /** Span ids backing this moment — the evidence trail. */
218
- evidenceSpanIds: string[];
219
- /** 1 (noise) … 5 (pivotal: failures, edits). Drives selection + duration. */
220
- importance: 1 | 2 | 3 | 4 | 5;
221
- startTs: number;
222
- endTs: number;
90
+ kind: SemanticKind;
91
+ /** One-line headline (the ticket/scene title). */
92
+ title: string;
93
+ /** Short human summary. */
94
+ summary: string;
95
+ /** The modality payload to show for this moment. */
96
+ visual: SceneVisual;
97
+ /** Span ids backing this moment — the evidence trail. */
98
+ evidenceSpanIds: string[];
99
+ /** 1 (noise) … 5 (pivotal: failures, edits). Drives selection + duration. */
100
+ importance: 1 | 2 | 3 | 4 | 5;
101
+ startTs: number;
102
+ endTs: number;
223
103
  }
224
104
  interface ReduceOptions {
225
- /** Collapse adjacent same-kind moments into one. Default true. */
226
- collapseAdjacent?: boolean;
105
+ /** Collapse adjacent same-kind moments into one. Default true. */
106
+ collapseAdjacent?: boolean;
227
107
  }
228
108
  /**
229
109
  * Reduce a span tree into the meaningful moments a viewer cares about. The
@@ -236,24 +116,24 @@ interface ReduceOptions {
236
116
  declare function reduceToSemanticEvents(spans: readonly Span[], opts?: ReduceOptions): SemanticEvent[];
237
117
  type SceneType = 'title_card' | 'prompt' | 'reply' | 'reasoning' | 'terminal' | 'file' | 'diff' | 'search' | 'browser' | 'computer' | 'api' | 'tool' | 'eval' | 'error' | 'summary';
238
118
  interface Scene {
239
- sceneType: SceneType;
240
- title: string;
241
- narration: string;
242
- /** The modality payload the renderer shows. */
243
- visual: SceneVisual;
244
- durationMs: number;
245
- evidenceSpanIds: string[];
119
+ sceneType: SceneType;
120
+ title: string;
121
+ narration: string;
122
+ /** The modality payload the renderer shows. */
123
+ visual: SceneVisual;
124
+ durationMs: number;
125
+ evidenceSpanIds: string[];
246
126
  }
247
127
  interface Storyboard {
248
- title: string;
249
- totalMs: number;
250
- scenes: Scene[];
128
+ title: string;
129
+ totalMs: number;
130
+ scenes: Scene[];
251
131
  }
252
132
  interface CompileOptions {
253
- /** Headline for the title card. Default "Agent run". */
254
- title?: string;
255
- /** Max action scenes between the title + summary cards. Default 16. */
256
- maxScenes?: number;
133
+ /** Headline for the title card. Default "Agent run". */
134
+ title?: string;
135
+ /** Max action scenes between the title + summary cards. Default 16. */
136
+ maxScenes?: number;
257
137
  }
258
138
  /**
259
139
  * Select the moments worth showing and lay them out as scenes. Every failure
@@ -264,8 +144,8 @@ declare function compileStoryboard(events: readonly SemanticEvent[], opts?: Comp
264
144
  /** Render the storyboard as a Markdown timeline — the human-readable shot list. */
265
145
  declare function renderStoryboardMarkdown(storyboard: Storyboard): string;
266
146
  interface HtmlRenderOptions {
267
- /** Document title + on-page heading. Defaults to the storyboard title. */
268
- title?: string;
147
+ /** Document title + on-page heading. Defaults to the storyboard title. */
148
+ title?: string;
269
149
  }
270
150
  /**
271
151
  * Render the storyboard as a self-contained, auto-playing HTML replay — one
@@ -275,5 +155,6 @@ interface HtmlRenderOptions {
275
155
  * run produces; an MP4 is the same storyboard fed to Remotion in a consumer.
276
156
  */
277
157
  declare function renderStoryboardHtml(storyboard: Storyboard, opts?: HtmlRenderOptions): string;
278
-
279
- export { type CodeAnimationOptions, type CodeEdit, type CompileOptions, type HtmlRenderOptions, type ReduceOptions, type Scene, type SceneType, type SceneVisual, type SemanticEvent, type SemanticKind, type Storyboard, codeEditFromSpan, codeEditsForStoryboard, compileStoryboard, editAnimationText, extractCodeEdits, reduceToSemanticEvents, renderCodeAnimationHtml, renderStoryboardHtml, renderStoryboardMarkdown };
158
+ //#endregion
159
+ export { CodeAnimationOptions, CodeEdit, CompileOptions, HtmlRenderOptions, ReduceOptions, Scene, SceneType, SceneVisual, SemanticEvent, SemanticKind, Storyboard, codeEditFromSpan, codeEditsForStoryboard, compileStoryboard, editAnimationText, extractCodeEdits, reduceToSemanticEvents, renderCodeAnimationHtml, renderStoryboardHtml, renderStoryboardMarkdown };
160
+ //# sourceMappingURL=index.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"index.d.ts","names":[],"sources":["../../src/storyboard/code-edit.ts","../../src/storyboard/index.ts"],"mappings":";;;UAuBiB;;EAEf;EACA;EACA;;EAEA;;;EAGA;EACA;EACA;EACA;EACA;;;iBA6Dc,iBAAiB,MAAM,OAAO;;iBAwB9B,iBAAiB,gBAAgB,SAAS;;;iBAS1C,uBACd,YAAY,YACZ,gBAAgB,SACf;EAAQ,OAAO;EAAO,MAAM;;;;;iBAgBf,kBAAkB,MAAM;UAiBvB;;EAEf;;EAEA;;EAEA;;;;;;;;;iBAUc,wBACd,iBAAiB,kBAAkB,YACnC,OAAM;;;;KCjKI;;;KAmBA;EACN;EAAoB;EAAa;;EACjC;EAAiB;EAAc;EAAiB;;EAChD;EAAc;EAAc;;EAC5B;EAAc;EAAc;;EAC5B;EAAkB;EAAiB;;EAEnC;EACA;EACA;EACA;EACA;EACA;;EAEA;EAAe;;EACf;;UAEW;EACf,MAAM;;EAEN;;EAEA;;EAEA,QAAQ;;EAER;;EAEA;EACA;EACA;;UAGe;;EAEf;;;;;;;;;;iBAiTc,uBACd,gBAAgB,QAChB,OAAM,gBACL;KA8CS;UAiBK;EACf,WAAW;EACX;EACA;;EAEA,QAAQ;EACR;EACA;;UAGe;EACf;EACA;EACA,QAAQ;;UAGO;;EAEf;;EAEA;;;;;;;iBAkCc,kBACd,iBAAiB,iBACjB,OAAM,iBACL;;iBA4Ja,yBAAyB,YAAY;UAwBpC;;EAEf;;;;;;;;;iBAUc,qBAAqB,YAAY,YAAY,OAAM"}