@tangle-network/agent-eval 0.215.1 → 0.216.1

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 (400) hide show
  1. package/CHANGELOG.md +51 -0
  2. package/dist/adapters/http.d.ts +3 -3
  3. package/dist/{agent-profile-DmzL8KaJ.d.ts → agent-profile-VFcQncbW.d.ts} +5 -5
  4. package/dist/{agent-profile-DmzL8KaJ.d.ts.map → agent-profile-VFcQncbW.d.ts.map} +1 -1
  5. package/dist/{agent-profile-cell-BYc0bxsL.js → agent-profile-cell-BK8mlqlq.js} +3 -3
  6. package/dist/{agent-profile-cell-BYc0bxsL.js.map → agent-profile-cell-BK8mlqlq.js.map} +1 -1
  7. package/dist/agreement-irr-BEWLIF3_.js +186 -0
  8. package/dist/agreement-irr-BEWLIF3_.js.map +1 -0
  9. package/dist/analyst/index.d.ts +487 -14
  10. package/dist/analyst/index.d.ts.map +1 -1
  11. package/dist/analyst/index.js +1547 -12
  12. package/dist/analyst/index.js.map +1 -1
  13. package/dist/{analyze-runs-3xpby3p2.js → analyze-runs-C84irm4i.js} +608 -604
  14. package/dist/analyze-runs-C84irm4i.js.map +1 -0
  15. package/dist/{analyze-runs-BhFLOztS.d.ts → analyze-runs-D68fd-su.d.ts} +17 -15
  16. package/dist/analyze-runs-D68fd-su.d.ts.map +1 -0
  17. package/dist/{baseline-Dm8966du.js → baseline-Udi_PLuT.js} +5 -5
  18. package/dist/{baseline-Dm8966du.js.map → baseline-Udi_PLuT.js.map} +1 -1
  19. package/dist/{benchmark-BywPeAUo.js → benchmark-CXmVVHH2.js} +2 -2
  20. package/dist/{benchmark-BywPeAUo.js.map → benchmark-CXmVVHH2.js.map} +1 -1
  21. package/dist/{benchmark-command-CgpGi7vY.js → benchmark-command-By0vh8q7.js} +71 -70
  22. package/dist/benchmark-command-By0vh8q7.js.map +1 -0
  23. package/dist/benchmarks/index.d.ts +5 -5
  24. package/dist/benchmarks/index.js +3 -3
  25. package/dist/builder-eval/index.js +2 -2
  26. package/dist/campaign/index.d.ts +7 -7
  27. package/dist/campaign/index.js +14 -14
  28. package/dist/{campaign-CpnsCgfz.js → campaign-BudZ70zo.js} +245 -247
  29. package/dist/campaign-BudZ70zo.js.map +1 -0
  30. package/dist/{campaign-evidence-OcUciD-N.js → campaign-evidence-Clrn4hBV.js} +8 -8
  31. package/dist/campaign-evidence-Clrn4hBV.js.map +1 -0
  32. package/dist/{canonical-BsZVgBhW.js → canonical-B2RqLofv.js} +2 -2
  33. package/dist/{canonical-BsZVgBhW.js.map → canonical-B2RqLofv.js.map} +1 -1
  34. package/dist/{chat-client-DthPOYGd.js → chat-client-l-MVDlDk.js} +2 -2
  35. package/dist/{chat-client-DthPOYGd.js.map → chat-client-l-MVDlDk.js.map} +1 -1
  36. package/dist/{chat-json-call-BrKsyspg.js → chat-json-call-puMArmtW.js} +2 -2
  37. package/dist/{chat-json-call-BrKsyspg.js.map → chat-json-call-puMArmtW.js.map} +1 -1
  38. package/dist/{claim-BXcZw-jo.js → claim-q4HQhK1G.js} +4 -4
  39. package/dist/claim-q4HQhK1G.js.map +1 -0
  40. package/dist/cli.js +9 -9
  41. package/dist/cli.js.map +1 -1
  42. package/dist/cluster-bootstrap-qcUWYFl-.js +66 -0
  43. package/dist/cluster-bootstrap-qcUWYFl-.js.map +1 -0
  44. package/dist/contract/index.d.ts +11 -11
  45. package/dist/contract/index.js +1023 -1079
  46. package/dist/contract/index.js.map +1 -1
  47. package/dist/{cost-ledger-D2FnJPBF.js → cost-ledger-BP8V4kUQ.js} +835 -816
  48. package/dist/cost-ledger-BP8V4kUQ.js.map +1 -0
  49. package/dist/{cost-ledger-CO7UyEcQ.d.ts → cost-ledger-Jk62OBfF.d.ts} +12 -47
  50. package/dist/cost-ledger-Jk62OBfF.d.ts.map +1 -0
  51. package/dist/{counterfactual-DtMWvn7D.js → counterfactual-Dn3veueY.js} +2 -2
  52. package/dist/{counterfactual-DtMWvn7D.js.map → counterfactual-Dn3veueY.js.map} +1 -1
  53. package/dist/{default-registry-KP4-Tolz.js → default-registry-COhk9g36.js} +744 -759
  54. package/dist/default-registry-COhk9g36.js.map +1 -0
  55. package/dist/{default-registry-B4B7GMvI.d.ts → default-registry-DzHiGruW.d.ts} +6 -6
  56. package/dist/{default-registry-B4B7GMvI.d.ts.map → default-registry-DzHiGruW.d.ts.map} +1 -1
  57. package/dist/{define-agent-eval-BdfnqXeP.js → define-agent-eval-CNTPzIkp.js} +10 -10
  58. package/dist/define-agent-eval-CNTPzIkp.js.map +1 -0
  59. package/dist/{define-agent-eval-DfXvNho5.d.ts → define-agent-eval-WSck0CZJ.d.ts} +8 -8
  60. package/dist/{define-agent-eval-DfXvNho5.d.ts.map → define-agent-eval-WSck0CZJ.d.ts.map} +1 -1
  61. package/dist/{descriptive-CnU8Yq29.js → descriptive-DboxPZrG.js} +2 -2
  62. package/dist/{descriptive-CnU8Yq29.js.map → descriptive-DboxPZrG.js.map} +1 -1
  63. package/dist/diagnosis/index.d.ts +3 -3
  64. package/dist/diagnosis/index.js +4 -4
  65. package/dist/{dspy-rlm-engine-DZfHM7kT.js → dspy-rlm-engine-LRCu61m5.js} +3 -3
  66. package/dist/{dspy-rlm-engine-DZfHM7kT.js.map → dspy-rlm-engine-LRCu61m5.js.map} +1 -1
  67. package/dist/{effect-sizes-DiH8MGOH.js → effect-sizes-BhzEpXMc.js} +2 -2
  68. package/dist/{effect-sizes-DiH8MGOH.js.map → effect-sizes-BhzEpXMc.js.map} +1 -1
  69. package/dist/{engine-C37HS2kr.d.ts → engine-3A_4jOlI.d.ts} +5 -5
  70. package/dist/{engine-C37HS2kr.d.ts.map → engine-3A_4jOlI.d.ts.map} +1 -1
  71. package/dist/{equal-terms-Blx9Zk5I.js → equal-terms-BtdXayCJ.js} +2 -2
  72. package/dist/{equal-terms-Blx9Zk5I.js.map → equal-terms-BtdXayCJ.js.map} +1 -1
  73. package/dist/{errors-Dngq5h35.js → errors-DVZiqS7K.js} +6 -2
  74. package/dist/{errors-Dngq5h35.js.map → errors-DVZiqS7K.js.map} +1 -1
  75. package/dist/{eval-campaign-4byTtqrf.js → eval-campaign-90jeaCAI.js} +6 -6
  76. package/dist/{eval-campaign-4byTtqrf.js.map → eval-campaign-90jeaCAI.js.map} +1 -1
  77. package/dist/evaluation-6Yoi74gM.d.ts +75 -0
  78. package/dist/evaluation-6Yoi74gM.d.ts.map +1 -0
  79. package/dist/evaluation.d.ts +2 -74
  80. package/dist/evaluation.js +1 -1
  81. package/dist/{exact-types-Dlb73wN4.d.ts → exact-types-ClugMJfI.d.ts} +2 -2
  82. package/dist/{exact-types-Dlb73wN4.d.ts.map → exact-types-ClugMJfI.d.ts.map} +1 -1
  83. package/dist/experiment/index.d.ts +404 -3
  84. package/dist/experiment/index.d.ts.map +1 -1
  85. package/dist/experiment/index.js +313 -19
  86. package/dist/experiment/index.js.map +1 -1
  87. package/dist/{external-optimizer-process-GngofQML.js → external-optimizer-process-BjMGUuds.js} +3 -3
  88. package/dist/{external-optimizer-process-GngofQML.js.map → external-optimizer-process-BjMGUuds.js.map} +1 -1
  89. package/dist/{external-optimizer-subprocess-BUxifb8g.js → external-optimizer-subprocess-BqW0hHcI.js} +791 -532
  90. package/dist/external-optimizer-subprocess-BqW0hHcI.js.map +1 -0
  91. package/dist/{feedback-trajectory-BLN_6fEs.d.ts → feedback-trajectory-C4EG0nNH.d.ts} +5 -8
  92. package/dist/feedback-trajectory-C4EG0nNH.d.ts.map +1 -0
  93. package/dist/fuzz.d.ts +1 -1
  94. package/dist/fuzz.js +3 -3
  95. package/dist/hosted/index.js +1 -1
  96. package/dist/{index-CPJJr98o.d.ts → index-BFRO6chs.d.ts} +3 -2
  97. package/dist/index-BFRO6chs.d.ts.map +1 -0
  98. package/dist/{index-Ch9VxCpK.d.ts → index-BO41eGj5.d.ts} +2 -2
  99. package/dist/{index-Ch9VxCpK.d.ts.map → index-BO41eGj5.d.ts.map} +1 -1
  100. package/dist/{index-B3WBm953.d.ts → index-CyszSg4O.d.ts} +165 -162
  101. package/dist/index-CyszSg4O.d.ts.map +1 -0
  102. package/dist/{index-B7m2uzJL.d.ts → index-DPcngtwt.d.ts} +30 -17
  103. package/dist/index-DPcngtwt.d.ts.map +1 -0
  104. package/dist/{index-tYLUvk_w.d.ts → index-DV-jSzxF.d.ts} +9 -9
  105. package/dist/{index-tYLUvk_w.d.ts.map → index-DV-jSzxF.d.ts.map} +1 -1
  106. package/dist/index.d.ts +1969 -1871
  107. package/dist/index.d.ts.map +1 -1
  108. package/dist/index.js +162 -201
  109. package/dist/index.js.map +1 -1
  110. package/dist/{integrity-BvnVARVl.js → integrity-BXOsAb3i.js} +2 -2
  111. package/dist/{integrity-BvnVARVl.js.map → integrity-BXOsAb3i.js.map} +1 -1
  112. package/dist/{integrity-CCAxGVK6.d.ts → integrity-DAMEDFaC.d.ts} +2 -2
  113. package/dist/{integrity-CCAxGVK6.d.ts.map → integrity-DAMEDFaC.d.ts.map} +1 -1
  114. package/dist/{integrity-BV1ZDf42.js → integrity-dLiDe2kD.js} +2 -5
  115. package/dist/integrity-dLiDe2kD.js.map +1 -0
  116. package/dist/{internal-CRHwS9iz.js → internal-Epl-YBZW.js} +3 -3
  117. package/dist/{internal-CRHwS9iz.js.map → internal-Epl-YBZW.js.map} +1 -1
  118. package/dist/interval-CELpye1n.js +219 -0
  119. package/dist/interval-CELpye1n.js.map +1 -0
  120. package/dist/interval-ohuOJurX.d.ts +47 -0
  121. package/dist/interval-ohuOJurX.d.ts.map +1 -0
  122. package/dist/{journal-DvKCvjty.js → journal-DfRz9kps.js} +2 -2
  123. package/dist/{journal-DvKCvjty.js.map → journal-DfRz9kps.js.map} +1 -1
  124. package/dist/{judge-calibration-N44V04Se.js → judge-calibration-BpCy5XQl.js} +3 -3
  125. package/dist/{judge-calibration-N44V04Se.js.map → judge-calibration-BpCy5XQl.js.map} +1 -1
  126. package/dist/{judge-gate-DNcUrRFA.d.ts → judge-gate-CE7CQAgt.d.ts} +3 -3
  127. package/dist/{judge-gate-DNcUrRFA.d.ts.map → judge-gate-CE7CQAgt.d.ts.map} +1 -1
  128. package/dist/{judge-gate-B1qYnkpi.js → judge-gate-iY9s4L44.js} +3 -3
  129. package/dist/judge-gate-iY9s4L44.js.map +1 -0
  130. package/dist/{kind-factory-AaEhUADH.js → kind-factory-BobI5Jy3.js} +5 -5
  131. package/dist/{kind-factory-AaEhUADH.js.map → kind-factory-BobI5Jy3.js.map} +1 -1
  132. package/dist/ledger-core/index.js +3 -3
  133. package/dist/{ledger-core-B2hahzrt.js → ledger-core-Gq7yp8lP.js} +4 -4
  134. package/dist/{ledger-core-B2hahzrt.js.map → ledger-core-Gq7yp8lP.js.map} +1 -1
  135. package/dist/{llm-client-DQveoxSM.js → llm-client-CyBAoYiX.js} +192 -219
  136. package/dist/llm-client-CyBAoYiX.js.map +1 -0
  137. package/dist/{llm-judge-Cd5s5Ei6.js → llm-judge-CG12SQJq.js} +499 -921
  138. package/dist/llm-judge-CG12SQJq.js.map +1 -0
  139. package/dist/matrix/index.d.ts +2 -2
  140. package/dist/matrix/index.js +1 -1
  141. package/dist/{matrix-DXVAJMLD.d.ts → matrix-DSrSxjjy.d.ts} +4 -4
  142. package/dist/{matrix-DXVAJMLD.d.ts.map → matrix-DSrSxjjy.d.ts.map} +1 -1
  143. package/dist/{matrix-_BV9-Lk5.js → matrix-Dm0BQW7U.js} +7 -11
  144. package/dist/matrix-Dm0BQW7U.js.map +1 -0
  145. package/dist/{paired-tests-jbwPdSz5.js → mean-DHoMiudD.js} +11 -5
  146. package/dist/mean-DHoMiudD.js.map +1 -0
  147. package/dist/meta-eval/index.d.ts +3 -3
  148. package/dist/meta-eval/index.js +9 -70
  149. package/dist/meta-eval/index.js.map +1 -1
  150. package/dist/mint-DNU8QPG4.d.ts +81 -0
  151. package/dist/mint-DNU8QPG4.d.ts.map +1 -0
  152. package/dist/{mint-Csor1BLJ.js → mint-QaD5cA0u.js} +3 -3
  153. package/dist/{mint-Csor1BLJ.js.map → mint-QaD5cA0u.js.map} +1 -1
  154. package/dist/{multiplicity-D9cjmStq.js → multiplicity-DQUFcbW4.js} +2 -2
  155. package/dist/{multiplicity-D9cjmStq.js.map → multiplicity-DQUFcbW4.js.map} +1 -1
  156. package/dist/multishot/golden/index.d.ts +1 -1
  157. package/dist/multishot/golden/index.d.ts.map +1 -1
  158. package/dist/multishot/golden/index.js +3 -6
  159. package/dist/multishot/golden/index.js.map +1 -1
  160. package/dist/multishot/index.d.ts +2 -2
  161. package/dist/multishot/index.js +1 -1
  162. package/dist/nonnegative-number-DvdMxg_f.js +9 -0
  163. package/dist/nonnegative-number-DvdMxg_f.js.map +1 -0
  164. package/dist/openapi.json +1 -1
  165. package/dist/{otel-spans-C8YIJWy1.js → otel-spans-B9tUXRHF.js} +3 -3
  166. package/dist/{otel-spans-C8YIJWy1.js.map → otel-spans-B9tUXRHF.js.map} +1 -1
  167. package/dist/{paired-arms-DzKn0TfY.js → paired-arms-CwMGJsrS.js} +5 -5
  168. package/dist/{paired-arms-DzKn0TfY.js.map → paired-arms-CwMGJsrS.js.map} +1 -1
  169. package/dist/{paired-binary-Q36bhn2R.js → paired-binary-CvpFHfKZ.js} +2 -2
  170. package/dist/{paired-binary-Q36bhn2R.js.map → paired-binary-CvpFHfKZ.js.map} +1 -1
  171. package/dist/{paired-promotion-decision-BN6IQTiu.js → paired-promotion-decision-DsAPYGCm.js} +3 -3
  172. package/dist/{paired-promotion-decision-BN6IQTiu.js.map → paired-promotion-decision-DsAPYGCm.js.map} +1 -1
  173. package/dist/{search-kernel-BztlERLv.js → pareto-Cm6ndrWN.js} +883 -861
  174. package/dist/pareto-Cm6ndrWN.js.map +1 -0
  175. package/dist/pipelines/index.d.ts +2 -2
  176. package/dist/pipelines/index.js +3 -3
  177. package/dist/{power-and-mde-fK5Qocwo.js → power-and-mde-Hj2ubb4B.js} +2 -2
  178. package/dist/{power-and-mde-fK5Qocwo.js.map → power-and-mde-Hj2ubb4B.js.map} +1 -1
  179. package/dist/predicate-FdKmElA_.js +38 -0
  180. package/dist/predicate-FdKmElA_.js.map +1 -0
  181. package/dist/{prime-protocol-DcAK9pC8.js → prime-protocol-CHKgA3a6.js} +2 -2
  182. package/dist/{prime-protocol-DcAK9pC8.js.map → prime-protocol-CHKgA3a6.js.map} +1 -1
  183. package/dist/{prime-protocol-BHia9vDz.d.ts → prime-protocol-HiB9HX7E.d.ts} +3 -3
  184. package/dist/{prime-protocol-BHia9vDz.d.ts.map → prime-protocol-HiB9HX7E.d.ts.map} +1 -1
  185. package/dist/{produced-state-BGp4BX8t.js → produced-state-BsDmnwxc.js} +8 -8
  186. package/dist/{produced-state-BGp4BX8t.js.map → produced-state-BsDmnwxc.js.map} +1 -1
  187. package/dist/profile-cell.js +1 -1
  188. package/dist/{release-confidence-BK2QCB5M.d.ts → promotion-gate-BawrdPYr.d.ts} +102 -102
  189. package/dist/promotion-gate-BawrdPYr.d.ts.map +1 -0
  190. package/dist/{promotion-policy-CWW35UWq.js → promotion-policy-KO57gbyV.js} +4 -4
  191. package/dist/{promotion-policy-CWW35UWq.js.map → promotion-policy-KO57gbyV.js.map} +1 -1
  192. package/dist/{proposal-findings-CXblsskH.js → proposal-findings-CY1xKn_a.js} +2 -36
  193. package/dist/proposal-findings-CY1xKn_a.js.map +1 -0
  194. package/dist/{query-B1D1qp2Z.js → query-BUCZCK7F.js} +3 -3
  195. package/dist/{query-B1D1qp2Z.js.map → query-BUCZCK7F.js.map} +1 -1
  196. package/dist/{random-Dn5fPWkt.js → random-ByUSnV9w.js} +2 -2
  197. package/dist/{random-Dn5fPWkt.js.map → random-ByUSnV9w.js.map} +1 -1
  198. package/dist/{raw-provider-sink-rB2DYxr6.js → raw-provider-sink-DoHz4JNG.js} +2 -2
  199. package/dist/{raw-provider-sink-rB2DYxr6.js.map → raw-provider-sink-DoHz4JNG.js.map} +1 -1
  200. package/dist/{redact-Dk1ZdJEm.js → redact-B5xXKqQ3.js} +97 -98
  201. package/dist/redact-B5xXKqQ3.js.map +1 -0
  202. package/dist/{redact-Cveoq6fJ.d.ts → redact-mtCCcT8z.d.ts} +4 -20
  203. package/dist/redact-mtCCcT8z.d.ts.map +1 -0
  204. package/dist/{registry-CMYIaDgS.d.ts → registry-D7_hm2fK.d.ts} +11 -13
  205. package/dist/registry-D7_hm2fK.d.ts.map +1 -0
  206. package/dist/{release-confidence-Dbe9AS_X.js → release-confidence-CpuP65S5.js} +5 -5
  207. package/dist/{release-confidence-Dbe9AS_X.js.map → release-confidence-CpuP65S5.js.map} +1 -1
  208. package/dist/{report-command-CtIIsQKP.js → report-command-SC9nqb4M.js} +128 -177
  209. package/dist/report-command-SC9nqb4M.js.map +1 -0
  210. package/dist/reporting-html-Kvlc8oMt.js +100 -0
  211. package/dist/reporting-html-Kvlc8oMt.js.map +1 -0
  212. package/dist/reporting.d.ts +3 -3
  213. package/dist/reporting.js +6 -6
  214. package/dist/{researcher-f6mZhh58.d.ts → researcher-CEcfku-T.d.ts} +5 -5
  215. package/dist/{researcher-f6mZhh58.d.ts.map → researcher-CEcfku-T.d.ts.map} +1 -1
  216. package/dist/reward-CpCmAM8L.d.ts +78 -0
  217. package/dist/reward-CpCmAM8L.d.ts.map +1 -0
  218. package/dist/{reward-hacking-BFcQgkux.js → reward-hacking-E0d-WVcP.js} +2 -2
  219. package/dist/{reward-hacking-BFcQgkux.js.map → reward-hacking-E0d-WVcP.js.map} +1 -1
  220. package/dist/{reward-hacking-DHA7c1l8.d.ts → reward-hacking-RVdZ6Kkt.d.ts} +2 -2
  221. package/dist/{reward-hacking-DHA7c1l8.d.ts.map → reward-hacking-RVdZ6Kkt.d.ts.map} +1 -1
  222. package/dist/rl.d.ts +8 -8
  223. package/dist/rl.js +14 -14
  224. package/dist/rollout/index.d.ts +3 -2
  225. package/dist/rollout/index.js +2 -2
  226. package/dist/{rollout-D9bKKBnZ.js → rollout-DZV1t2Wg.js} +2 -2
  227. package/dist/{rollout-D9bKKBnZ.js.map → rollout-DZV1t2Wg.js.map} +1 -1
  228. package/dist/round-decimal-COrKglMh.js +10 -0
  229. package/dist/round-decimal-COrKglMh.js.map +1 -0
  230. package/dist/{rubric-predictive-validity-2SgHxoa5.d.ts → rubric-predictive-validity-C03EC7ZE.d.ts} +2 -2
  231. package/dist/{rubric-predictive-validity-2SgHxoa5.d.ts.map → rubric-predictive-validity-C03EC7ZE.d.ts.map} +1 -1
  232. package/dist/{rubric-predictive-validity-BsgkBfT0.js → rubric-predictive-validity-DaFW70y9.js} +3 -3
  233. package/dist/{rubric-predictive-validity-BsgkBfT0.js.map → rubric-predictive-validity-DaFW70y9.js.map} +1 -1
  234. package/dist/{run-record-DlTZ_lqe.js → run-record-Boc5IUTp.js} +4 -4
  235. package/dist/{run-record-DlTZ_lqe.js.map → run-record-Boc5IUTp.js.map} +1 -1
  236. package/dist/{run-record-D1Hw-Jml.d.ts → run-record-C1E24G2A.d.ts} +2 -2
  237. package/dist/{run-record-D1Hw-Jml.d.ts.map → run-record-C1E24G2A.d.ts.map} +1 -1
  238. package/dist/{run-record-ES07wfz9.js → run-record-C5e7FjhJ.js} +3 -5
  239. package/dist/run-record-C5e7FjhJ.js.map +1 -0
  240. package/dist/search/index.d.ts +46 -32
  241. package/dist/search/index.d.ts.map +1 -1
  242. package/dist/search/index.js +3 -3
  243. package/dist/{search-kernel-BqyBZ9Wm.d.ts → search-kernel-DyNEa6IU.d.ts} +14 -10
  244. package/dist/search-kernel-DyNEa6IU.d.ts.map +1 -0
  245. package/dist/{search-ledger-d7V8o9Lk.js → search-ledger-CVUGLDZp.js} +189 -199
  246. package/dist/search-ledger-CVUGLDZp.js.map +1 -0
  247. package/dist/{search-ledger-file-BqOo-bVo.js → search-ledger-file-DDQUgvCg.js} +4 -4
  248. package/dist/{search-ledger-file-BqOo-bVo.js.map → search-ledger-file-DDQUgvCg.js.map} +1 -1
  249. package/dist/search-ledger-types-p2xKfiDJ.d.ts.map +1 -1
  250. package/dist/{search-shipper-er4pK5_Y.js → search-shipper-BOSD46qv.js} +3 -3
  251. package/dist/{search-shipper-er4pK5_Y.js.map → search-shipper-BOSD46qv.js.map} +1 -1
  252. package/dist/{search-summary-CBvz3u_S.js → search-summary-BQCSWf1g.js} +5 -5
  253. package/dist/{search-summary-CBvz3u_S.js.map → search-summary-BQCSWf1g.js.map} +1 -1
  254. package/dist/{semantic-concept-judge-BPqbwIVs.js → semantic-concept-judge-BQF8te0T.js} +12 -14
  255. package/dist/semantic-concept-judge-BQF8te0T.js.map +1 -0
  256. package/dist/{sequential-D-1PPCmM.js → sequential-Bf_pbE50.js} +5 -5
  257. package/dist/{sequential-D-1PPCmM.js.map → sequential-Bf_pbE50.js.map} +1 -1
  258. package/dist/{sequential-eprocess-D1jKoihe.js → sequential-eprocess-DM_qENCl.js} +2 -2
  259. package/dist/{sequential-eprocess-D1jKoihe.js.map → sequential-eprocess-DM_qENCl.js.map} +1 -1
  260. package/dist/{server-DbcC7xMl.js → server-e9oWO_bN.js} +4 -4
  261. package/dist/{server-DbcC7xMl.js.map → server-e9oWO_bN.js.map} +1 -1
  262. package/dist/{single-run-lock-fVu3MZWH.js → single-run-lock-CV9Eyd_q.js} +2 -2
  263. package/dist/{single-run-lock-fVu3MZWH.js.map → single-run-lock-CV9Eyd_q.js.map} +1 -1
  264. package/dist/{skillopt-optimization-method-Dcepr6QL.js → skillopt-optimization-method-BiwA8nfX.js} +427 -12
  265. package/dist/skillopt-optimization-method-BiwA8nfX.js.map +1 -0
  266. package/dist/{spans-D9_ehFuk.d.ts → spans-BHAUXMGN.d.ts} +2 -2
  267. package/dist/{spans-D9_ehFuk.d.ts.map → spans-BHAUXMGN.d.ts.map} +1 -1
  268. package/dist/{statistical-heldout-1Gbmex7v.js → statistical-heldout-42NlCNDe.js} +3 -3
  269. package/dist/{statistical-heldout-1Gbmex7v.js.map → statistical-heldout-42NlCNDe.js.map} +1 -1
  270. package/dist/{statistical-heldout-CJTj2EoS.d.ts → statistical-heldout-CYpGaoDH.d.ts} +3 -3
  271. package/dist/{statistical-heldout-CJTj2EoS.d.ts.map → statistical-heldout-CYpGaoDH.d.ts.map} +1 -1
  272. package/dist/{store-otlp-DWVhUgEj.js → store-otlp-DJ9dtA5x.js} +135 -132
  273. package/dist/store-otlp-DJ9dtA5x.js.map +1 -0
  274. package/dist/{store-tool-spans-BNICrhsb.js → store-tool-spans-CIX5arAG.js} +7 -7
  275. package/dist/{store-tool-spans-BNICrhsb.js.map → store-tool-spans-CIX5arAG.js.map} +1 -1
  276. package/dist/{store-tool-spans-BNzls3OD.d.ts → store-tool-spans-DdCSPa7C.d.ts} +162 -161
  277. package/dist/store-tool-spans-DdCSPa7C.d.ts.map +1 -0
  278. package/dist/storyboard/index.d.ts.map +1 -1
  279. package/dist/storyboard/index.js +20 -3
  280. package/dist/storyboard/index.js.map +1 -1
  281. package/dist/{student-t-BQeKJx2N.js → student-t-DVPW7CdL.js} +2 -2
  282. package/dist/{student-t-BQeKJx2N.js.map → student-t-DVPW7CdL.js.map} +1 -1
  283. package/dist/{summary-report-DtdRucM0.js → summary-report-V_JACR4Y.js} +526 -614
  284. package/dist/summary-report-V_JACR4Y.js.map +1 -0
  285. package/dist/{summary-report-C-CtItCN.d.ts → summary-report-bdFDXq0b.d.ts} +10 -5
  286. package/dist/summary-report-bdFDXq0b.d.ts.map +1 -0
  287. package/dist/supervisor-run/index.d.ts +154 -150
  288. package/dist/supervisor-run/index.d.ts.map +1 -1
  289. package/dist/supervisor-run/index.js +65 -3
  290. package/dist/supervisor-run/index.js.map +1 -1
  291. package/dist/systemone-BomhHr2c.d.ts +135 -0
  292. package/dist/systemone-BomhHr2c.d.ts.map +1 -0
  293. package/dist/systemone-protocol.js +1 -1
  294. package/dist/{systemone-review-C4dZgrgN.d.ts → systemone-review-CUVcQJu6.d.ts} +2 -2
  295. package/dist/{systemone-review-C4dZgrgN.d.ts.map → systemone-review-CUVcQJu6.d.ts.map} +1 -1
  296. package/dist/{systemone-review-DTClwgmE.js → systemone-review-D7sXwsqq.js} +3 -3
  297. package/dist/{systemone-review-DTClwgmE.js.map → systemone-review-D7sXwsqq.js.map} +1 -1
  298. package/dist/systemone.d.ts +4 -134
  299. package/dist/systemone.js +4 -4
  300. package/dist/{task-failure-attributes-DngaLdE4.js → task-failure-attributes-hEmASjix.js} +2 -2
  301. package/dist/{task-failure-attributes-DngaLdE4.js.map → task-failure-attributes-hEmASjix.js.map} +1 -1
  302. package/dist/{terminal-record-BHocwFw1.js → terminal-record-DZyQZ3pz.js} +14 -10
  303. package/dist/terminal-record-DZyQZ3pz.js.map +1 -0
  304. package/dist/{tool-groups-BKRrtqy4.d.ts → tool-groups-C3NEcdCM.d.ts} +3 -3
  305. package/dist/tool-groups-C3NEcdCM.d.ts.map +1 -0
  306. package/dist/{tool-waste-Wi4g6Ekb.d.ts → tool-waste-BKgvd-TP.d.ts} +30 -30
  307. package/dist/tool-waste-BKgvd-TP.d.ts.map +1 -0
  308. package/dist/{tool-waste-BYt5rdt8.js → tool-waste-BlesqTqW.js} +5 -158
  309. package/dist/tool-waste-BlesqTqW.js.map +1 -0
  310. package/dist/{trace-questions-CLx8vt6t.d.ts → trace-questions-CM2CAocU.d.ts} +4 -4
  311. package/dist/{trace-questions-CLx8vt6t.d.ts.map → trace-questions-CM2CAocU.d.ts.map} +1 -1
  312. package/dist/trace-repair/index.d.ts +3 -3
  313. package/dist/trace-repair/index.d.ts.map +1 -1
  314. package/dist/trace-repair/index.js +7 -13
  315. package/dist/trace-repair/index.js.map +1 -1
  316. package/dist/traces.d.ts +76 -58
  317. package/dist/traces.d.ts.map +1 -1
  318. package/dist/traces.js +9 -9
  319. package/dist/trajectory-replay/index.js +3 -3
  320. package/dist/{tree-rXLW3M_4.js → tree-C1XYeXvy.js} +2277 -2339
  321. package/dist/tree-C1XYeXvy.js.map +1 -0
  322. package/dist/{types-DjOMwGY_.d.ts → types-BqkVmHfv.d.ts} +7 -1
  323. package/dist/types-BqkVmHfv.d.ts.map +1 -0
  324. package/dist/{types-DBZ7keha.d.ts → types-C02qtnG8.d.ts} +4 -46
  325. package/dist/types-C02qtnG8.d.ts.map +1 -0
  326. package/dist/{types-DHb5TI36.d.ts → types-C1tnkEgM.d.ts} +347 -6
  327. package/dist/types-C1tnkEgM.d.ts.map +1 -0
  328. package/dist/{types-DinHtRJp.d.ts → types-Cbe2fWBa.d.ts} +3 -3
  329. package/dist/{types-DinHtRJp.d.ts.map → types-Cbe2fWBa.d.ts.map} +1 -1
  330. package/dist/{types-CGEmVYOk.d.ts → types-Cq8mSbq8.d.ts} +78 -72
  331. package/dist/types-Cq8mSbq8.d.ts.map +1 -0
  332. package/dist/{types-CW5b90JH.js → types-IXD2HJzu.js} +16 -3
  333. package/dist/types-IXD2HJzu.js.map +1 -0
  334. package/dist/{verdict-DhPoXApk.js → verdict-Dl0tiT1X.js} +2 -2
  335. package/dist/{verdict-DhPoXApk.js.map → verdict-Dl0tiT1X.js.map} +1 -1
  336. package/dist/{verdict-cache-BQEsVy9n.js → verdict-cache-Bm8EKL2s.js} +2 -2
  337. package/dist/{verdict-cache-BQEsVy9n.js.map → verdict-cache-Bm8EKL2s.js.map} +1 -1
  338. package/dist/wire/index.d.ts +3 -3
  339. package/dist/wire/index.js +1 -1
  340. package/docs/concepts.md +1 -0
  341. package/docs/design/mlbenchmarks-book-review.md +1 -1
  342. package/docs/design/mlbenchmarks-review/probes.mts +23 -31
  343. package/docs/experiment.md +1 -1
  344. package/docs/finding-layers.md +69 -0
  345. package/package.json +1 -1
  346. package/dist/analyze-runs-3xpby3p2.js.map +0 -1
  347. package/dist/analyze-runs-BhFLOztS.d.ts.map +0 -1
  348. package/dist/ast-DJ6Bc7ab.js +0 -558
  349. package/dist/ast-DJ6Bc7ab.js.map +0 -1
  350. package/dist/ast-hI-vjW6J.d.ts +0 -457
  351. package/dist/ast-hI-vjW6J.d.ts.map +0 -1
  352. package/dist/benchmark-command-CgpGi7vY.js.map +0 -1
  353. package/dist/campaign-CpnsCgfz.js.map +0 -1
  354. package/dist/campaign-evidence-OcUciD-N.js.map +0 -1
  355. package/dist/claim-BXcZw-jo.js.map +0 -1
  356. package/dist/cost-ledger-CO7UyEcQ.d.ts.map +0 -1
  357. package/dist/cost-ledger-D2FnJPBF.js.map +0 -1
  358. package/dist/default-registry-KP4-Tolz.js.map +0 -1
  359. package/dist/define-agent-eval-BdfnqXeP.js.map +0 -1
  360. package/dist/evaluation.d.ts.map +0 -1
  361. package/dist/external-optimizer-subprocess-BUxifb8g.js.map +0 -1
  362. package/dist/feedback-trajectory-BLN_6fEs.d.ts.map +0 -1
  363. package/dist/index-B3WBm953.d.ts.map +0 -1
  364. package/dist/index-B7m2uzJL.d.ts.map +0 -1
  365. package/dist/index-CPJJr98o.d.ts.map +0 -1
  366. package/dist/integrity-BV1ZDf42.js.map +0 -1
  367. package/dist/judge-gate-B1qYnkpi.js.map +0 -1
  368. package/dist/llm-client-DQveoxSM.js.map +0 -1
  369. package/dist/llm-judge-Cd5s5Ei6.js.map +0 -1
  370. package/dist/matrix-_BV9-Lk5.js.map +0 -1
  371. package/dist/mint-DCGoBYhB.d.ts +0 -155
  372. package/dist/mint-DCGoBYhB.d.ts.map +0 -1
  373. package/dist/paired-tests-jbwPdSz5.js.map +0 -1
  374. package/dist/proposal-findings-CXblsskH.js.map +0 -1
  375. package/dist/redact-Cveoq6fJ.d.ts.map +0 -1
  376. package/dist/redact-Dk1ZdJEm.js.map +0 -1
  377. package/dist/registry-CMYIaDgS.d.ts.map +0 -1
  378. package/dist/release-confidence-BK2QCB5M.d.ts.map +0 -1
  379. package/dist/report-command-CtIIsQKP.js.map +0 -1
  380. package/dist/run-record-ES07wfz9.js.map +0 -1
  381. package/dist/search-kernel-BqyBZ9Wm.d.ts.map +0 -1
  382. package/dist/search-kernel-BztlERLv.js.map +0 -1
  383. package/dist/search-ledger-d7V8o9Lk.js.map +0 -1
  384. package/dist/semantic-concept-judge-BPqbwIVs.js.map +0 -1
  385. package/dist/skillopt-optimization-method-Dcepr6QL.js.map +0 -1
  386. package/dist/store-otlp-DWVhUgEj.js.map +0 -1
  387. package/dist/store-tool-spans-BNzls3OD.d.ts.map +0 -1
  388. package/dist/summary-report-C-CtItCN.d.ts.map +0 -1
  389. package/dist/summary-report-DtdRucM0.js.map +0 -1
  390. package/dist/systemone.d.ts.map +0 -1
  391. package/dist/terminal-record-BHocwFw1.js.map +0 -1
  392. package/dist/tool-groups-BKRrtqy4.d.ts.map +0 -1
  393. package/dist/tool-waste-BYt5rdt8.js.map +0 -1
  394. package/dist/tool-waste-Wi4g6Ekb.d.ts.map +0 -1
  395. package/dist/tree-rXLW3M_4.js.map +0 -1
  396. package/dist/types-CGEmVYOk.d.ts.map +0 -1
  397. package/dist/types-CW5b90JH.js.map +0 -1
  398. package/dist/types-DBZ7keha.d.ts.map +0 -1
  399. package/dist/types-DHb5TI36.d.ts.map +0 -1
  400. package/dist/types-DjOMwGY_.d.ts.map +0 -1
@@ -1,149 +1,13 @@
1
+ import { a as CostLedgerHandle } from "./cost-ledger-Jk62OBfF.js";
1
2
  import { r as CaptureIntegrityError } from "./errors-DEE6u6ot.js";
2
- import { c as CostLedgerHandle } from "./cost-ledger-CO7UyEcQ.js";
3
3
  import { C as ToolSpan } from "./schema-2g7trUwH.js";
4
- import { d as RunTokenUsage } from "./run-record-D1Hw-Jml.js";
5
- import { B as TraceAnalystSpan, E as SpanSourceReader, F as SearchTraceResult, K as ViewTraceResult, N as QueryTracesPage, O as TraceAnalysisStore, P as SearchSpanResult, W as ViewSpansResult, f as AnalystUsageReceipt, i as AnalystFinding, j as DatasetOverview, k as TraceAnalysisStoreContext, z as TraceAnalystFilters } from "./types-DHb5TI36.js";
6
- import { et as ProviderRedactor, rt as RawProviderSink } from "./types-DBZ7keha.js";
4
+ import { d as RunTokenUsage } from "./run-record-C1E24G2A.js";
5
+ import { $ as DatasetOverview, J as SpanSourceReader, X as TraceAnalysisStore, Z as TraceAnalysisStoreContext, a as AnalystFinding, ct as TraceAnalystSpan, ft as ViewSpansResult, m as AnalystUsageReceipt, mt as ViewTraceResult, nt as SearchSpanResult, rt as SearchTraceResult, st as TraceAnalystFilters, tt as QueryTracesPage } from "./types-C1tnkEgM.js";
6
+ import { X as RawProviderSink, q as ProviderRedactor } from "./types-C02qtnG8.js";
7
7
  import { i as SearchArtifactRef } from "./search-ledger-types-p2xKfiDJ.js";
8
8
  import { s as TraceStore } from "./store-C4avFjIi.js";
9
- import { a as TraceAnalystLimits, i as TraceAnalysisEngineResult, n as TraceAnalysisEngine } from "./engine-C37HS2kr.js";
10
- import { t as TraceToolGroupName } from "./tool-groups-BKRrtqy4.js";
11
- //#region src/trace/extract-usage.d.ts
12
- type ExtractedUsage = RunTokenUsage;
13
- type SseUsageMode = 'cumulative' | 'delta';
14
- interface ExtractUsageFromSseOptions {
15
- /** Provider usage events are cumulative snapshots unless explicitly marked as deltas. */
16
- mode?: SseUsageMode;
17
- }
18
- /**
19
- * Pull `{ input, output, cached?, cacheWrite? }` from a parsed response
20
- * body. Accepts a top-level `usage` object (the common case) or a body that IS
21
- * the usage object. Returns null only when none of those categories is present.
22
- */
23
- declare function extractUsage(body: unknown): ExtractedUsage | null;
24
- /**
25
- * Extract token usage from a complete SSE response body using the shared SSE
26
- * frame parser. Cumulative snapshots are the fail-safe default because summing
27
- * them inflates billing; callers with explicit delta events opt into `mode: 'delta'`.
28
- */
29
- declare function extractUsageFromSse(text: string, options?: ExtractUsageFromSseOptions): ExtractedUsage | null;
30
- /**
31
- * Extract usage from an HTTP `Response` without consuming the caller's body:
32
- * clones, reads the text, and tries the JSON parser first, then the SSE
33
- * accumulator. Best-effort — returns null on any read/parse miss so a usage tee
34
- * never takes down the underlying call.
35
- */
36
- declare function extractUsageFromResponse(response: Response, sseOptions?: ExtractUsageFromSseOptions): Promise<ExtractedUsage | null>;
37
- //#endregion
38
- //#region src/trace/capture-fetch.d.ts
39
- interface CaptureFetchContext {
40
- /** Logical run id stamped on every captured event. Required — without it
41
- * the raw events can't be paired with their parent `Run`. */
42
- runId: string;
43
- /** Optional logical span id (enables span-level sink filtering). */
44
- spanId?: string;
45
- /** Resolved base URL (normalised, no trailing slash). Used for the event's
46
- * `baseUrl` and for endpoint-path extraction. */
47
- baseUrl: string;
48
- /** Model id the caller intends to invoke. Stamped on every event. */
49
- model: string;
50
- /** Provider override. When omitted, `providerFromBaseUrl(baseUrl)`. */
51
- provider?: string;
52
- }
53
- interface CaptureFetchOptions {
54
- /** Override the capture-time redactor. Default `defaultProviderRedactor`. */
55
- redactor?: ProviderRedactor;
56
- /** Cap on captured response-body bytes; beyond it the body is truncated and
57
- * `body_truncated` is added to `redactedFields`. Default 2 MiB. */
58
- responseBodyByteCap?: number;
59
- /** When true, a sink-write failure propagates to the caller. Default false
60
- * — capture is best-effort so a sink failure never kills the LLM call. */
61
- failClosed?: boolean;
62
- /**
63
- * Invoked with the token usage parsed off each successful response (JSON or
64
- * SSE), keyed by the captured context. Lets a caller fold usage → cost without
65
- * re-cloning the response themselves. Not called when the response carries no
66
- * usage. Best-effort: a throw here is swallowed (it never kills the LLM call)
67
- * unless `failClosed` is set.
68
- */
69
- onUsage?: (usage: ExtractedUsage, ctx: CaptureFetchContext) => void;
70
- /** How repeated SSE usage events should be reconciled. Default `cumulative`. */
71
- sseUsageMode?: SseUsageMode;
72
- }
73
- declare function captureFetchToRawSink(fetch: typeof globalThis.fetch, sink: RawProviderSink, ctx: CaptureFetchContext, opts?: CaptureFetchOptions): typeof globalThis.fetch;
74
- //#endregion
75
- //#region src/trace/otel.d.ts
76
- declare const OTEL_AGENT_EVAL_SCOPE: {
77
- name: string;
78
- version: string;
79
- };
80
- interface OtlpSpan {
81
- traceId: string;
82
- spanId: string;
83
- parentSpanId?: string;
84
- name: string;
85
- kind: number;
86
- startTimeUnixNano: string;
87
- endTimeUnixNano: string;
88
- attributes: Array<{
89
- key: string;
90
- value: {
91
- stringValue?: string;
92
- intValue?: string;
93
- doubleValue?: number;
94
- boolValue?: boolean;
95
- };
96
- }>;
97
- events?: Array<{
98
- timeUnixNano: string;
99
- name: string;
100
- attributes?: OtlpSpan['attributes'];
101
- }>;
102
- status?: {
103
- code: number;
104
- message?: string;
105
- };
106
- }
107
- interface OtlpResourceSpans {
108
- resource: {
109
- attributes: OtlpSpan['attributes'];
110
- };
111
- scopeSpans: Array<{
112
- scope: typeof OTEL_AGENT_EVAL_SCOPE;
113
- spans: OtlpSpan[];
114
- }>;
115
- }
116
- interface OtlpExport {
117
- resourceSpans: OtlpResourceSpans[];
118
- }
119
- /** Export a single run's spans + events in OTLP/JSON. */
120
- declare function exportRunAsOtlp(store: TraceStore, runId: string, resourceAttrs?: Record<string, string | number | boolean>): Promise<OtlpExport>;
121
- //#endregion
122
- //#region src/trace/otlp-flat.d.ts
123
- type OtlpStatusCode = 'STATUS_CODE_OK' | 'STATUS_CODE_ERROR' | 'STATUS_CODE_UNSET';
124
- interface OtlpFlatLine {
125
- trace_id: string;
126
- span_id: string;
127
- parent_span_id: string | null;
128
- name: string;
129
- kind: string;
130
- start_time: string;
131
- end_time: string;
132
- status: {
133
- code: OtlpStatusCode;
134
- message?: string;
135
- };
136
- resource: {
137
- attributes: Record<string, unknown>;
138
- };
139
- attributes: Record<string, unknown>;
140
- events?: Array<{
141
- name: string;
142
- timeUnixNano?: string;
143
- attributes?: Record<string, unknown>;
144
- }>;
145
- }
146
- //#endregion
9
+ import { a as TraceAnalystLimits, i as TraceAnalysisEngineResult, n as TraceAnalysisEngine } from "./engine-3A_4jOlI.js";
10
+ import { t as TraceToolGroupName } from "./tool-groups-C3NEcdCM.js";
147
11
  //#region src/trace-analyst/analyst.d.ts
148
12
  interface AnalyzeTracesInput {
149
13
  question: string;
@@ -300,7 +164,54 @@ declare function scoreTraceInsightReadiness(context: TraceInsightContext): Trace
300
164
  declare function defaultTraceInsightPanel(): TraceInsightPanelRole[];
301
165
  declare function buildTraceInsightPrompt(input: TraceInsightPromptInput): string;
302
166
  //#endregion
303
- //#region src/trace-analyst/store-otlp.d.ts
167
+ //#region src/trace/otel.d.ts
168
+ declare const OTEL_AGENT_EVAL_SCOPE: {
169
+ name: string;
170
+ version: string;
171
+ };
172
+ interface OtlpSpan {
173
+ traceId: string;
174
+ spanId: string;
175
+ parentSpanId?: string;
176
+ name: string;
177
+ kind: number;
178
+ startTimeUnixNano: string;
179
+ endTimeUnixNano: string;
180
+ attributes: Array<{
181
+ key: string;
182
+ value: {
183
+ stringValue?: string;
184
+ intValue?: string;
185
+ doubleValue?: number;
186
+ boolValue?: boolean;
187
+ };
188
+ }>;
189
+ events?: Array<{
190
+ timeUnixNano: string;
191
+ name: string;
192
+ attributes?: OtlpSpan['attributes'];
193
+ }>;
194
+ status?: {
195
+ code: number;
196
+ message?: string;
197
+ };
198
+ }
199
+ interface OtlpResourceSpans {
200
+ resource: {
201
+ attributes: OtlpSpan['attributes'];
202
+ };
203
+ scopeSpans: Array<{
204
+ scope: typeof OTEL_AGENT_EVAL_SCOPE;
205
+ spans: OtlpSpan[];
206
+ }>;
207
+ }
208
+ interface OtlpExport {
209
+ resourceSpans: OtlpResourceSpans[];
210
+ }
211
+ /** Export a single run's spans + events in OTLP/JSON. */
212
+ declare function exportRunAsOtlp(store: TraceStore, runId: string, resourceAttrs?: Record<string, string | number | boolean>): Promise<OtlpExport>;
213
+ //#endregion
214
+ //#region src/trace-analyst/store-otlp-buffered.d.ts
304
215
  interface ToolSpansToTraceAnalysisStoreOptions {
305
216
  /** Authorized original-record lookup, exposed only when supplied. */
306
217
  sourceReader?: SpanSourceReader;
@@ -313,21 +224,6 @@ interface ToolSpansToTraceAnalysisStoreOptions {
313
224
  /** Override the per-match text budget. */
314
225
  perMatchTextBudget?: number;
315
226
  }
316
- /** Budgets and authorized source access shared by canonical span-record stores. */
317
- type SpanRecordsToTraceAnalysisStoreOptions = ToolSpansToTraceAnalysisStoreOptions;
318
- type BufferedTraceStoreOptions = ToolSpansToTraceAnalysisStoreOptions;
319
- interface OtlpFileTraceStoreOptions extends BufferedTraceStoreOptions {
320
- /** Path to the OTLP-JSONL file. */
321
- path: string;
322
- /**
323
- * Hard ceiling on the trace file size in bytes. The store reads the
324
- * whole file into one Buffer and indexes it in memory, so an
325
- * unbounded file OOMs the process. Above this size the store fails
326
- * loud with `TraceFileTooLargeError` instead of degrading silently.
327
- * Default 256 MiB.
328
- */
329
- maxFileBytes?: number;
330
- }
331
227
  declare abstract class BufferedTraceStore implements TraceAnalysisStore {
332
228
  readonly readSpanSource?: SpanSourceReader;
333
229
  private readonly perAttributeViewBudget;
@@ -338,7 +234,7 @@ declare abstract class BufferedTraceStore implements TraceAnalysisStore {
338
234
  private indexTask?;
339
235
  private bufferValue?;
340
236
  private bufferTask?;
341
- constructor(opts: BufferedTraceStoreOptions);
237
+ constructor(opts: ToolSpansToTraceAnalysisStoreOptions);
342
238
  hasTrace(trace_id: string, context?: TraceAnalysisStoreContext): Promise<boolean>;
343
239
  hasSpans(opts: {
344
240
  trace_id: string;
@@ -386,6 +282,22 @@ declare abstract class BufferedTraceStore implements TraceAnalysisStore {
386
282
  private buildOversizedSummary;
387
283
  private scanSpanForMatches;
388
284
  }
285
+ //#endregion
286
+ //#region src/trace-analyst/store-otlp.d.ts
287
+ interface OtlpFileTraceStoreOptions extends ToolSpansToTraceAnalysisStoreOptions {
288
+ /** Path to the OTLP-JSONL file. */
289
+ path: string;
290
+ /**
291
+ * Hard ceiling on the trace file size in bytes. The store reads the
292
+ * whole file into one Buffer and indexes it in memory, so an
293
+ * unbounded file OOMs the process. Above this size the store fails
294
+ * loud with `TraceFileTooLargeError` instead of degrading silently.
295
+ * Default 256 MiB.
296
+ */
297
+ maxFileBytes?: number;
298
+ }
299
+ /** Budgets and authorized source access shared by canonical span-record stores. */
300
+ type SpanRecordsToTraceAnalysisStoreOptions = ToolSpansToTraceAnalysisStoreOptions;
389
301
  declare class OtlpFileTraceStore extends BufferedTraceStore {
390
302
  private readonly path;
391
303
  private readonly maxFileBytes;
@@ -404,6 +316,95 @@ declare function spanRecordsToTraceAnalysisStore(records: readonly TraceAnalystS
404
316
  /** Build the bounded trace store directly from OpenInference JSONL text. */
405
317
  declare function otlpTextToTraceAnalysisStore(text: string, opts?: ToolSpansToTraceAnalysisStoreOptions): TraceAnalysisStore;
406
318
  //#endregion
319
+ //#region src/trace/extract-usage.d.ts
320
+ type ExtractedUsage = RunTokenUsage;
321
+ type SseUsageMode = 'cumulative' | 'delta';
322
+ interface ExtractUsageFromSseOptions {
323
+ /** Provider usage events are cumulative snapshots unless explicitly marked as deltas. */
324
+ mode?: SseUsageMode;
325
+ }
326
+ /**
327
+ * Pull `{ input, output, cached?, cacheWrite? }` from a parsed response
328
+ * body. Accepts a top-level `usage` object (the common case) or a body that IS
329
+ * the usage object. Returns null only when none of those categories is present.
330
+ */
331
+ declare function extractUsage(body: unknown): ExtractedUsage | null;
332
+ /**
333
+ * Extract token usage from a complete SSE response body using the shared SSE
334
+ * frame parser. Cumulative snapshots are the fail-safe default because summing
335
+ * them inflates billing; callers with explicit delta events opt into `mode: 'delta'`.
336
+ */
337
+ declare function extractUsageFromSse(text: string, options?: ExtractUsageFromSseOptions): ExtractedUsage | null;
338
+ /**
339
+ * Extract usage from an HTTP `Response` without consuming the caller's body:
340
+ * clones, reads the text, and tries the JSON parser first, then the SSE
341
+ * accumulator. Best-effort — returns null on any read/parse miss so a usage tee
342
+ * never takes down the underlying call.
343
+ */
344
+ declare function extractUsageFromResponse(response: Response, sseOptions?: ExtractUsageFromSseOptions): Promise<ExtractedUsage | null>;
345
+ //#endregion
346
+ //#region src/trace/otlp-flat.d.ts
347
+ type OtlpStatusCode = 'STATUS_CODE_OK' | 'STATUS_CODE_ERROR' | 'STATUS_CODE_UNSET';
348
+ interface OtlpFlatLine {
349
+ trace_id: string;
350
+ span_id: string;
351
+ parent_span_id: string | null;
352
+ name: string;
353
+ kind: string;
354
+ start_time: string;
355
+ end_time: string;
356
+ status: {
357
+ code: OtlpStatusCode;
358
+ message?: string;
359
+ };
360
+ resource: {
361
+ attributes: Record<string, unknown>;
362
+ };
363
+ attributes: Record<string, unknown>;
364
+ events?: Array<{
365
+ name: string;
366
+ timeUnixNano?: string;
367
+ attributes?: Record<string, unknown>;
368
+ }>;
369
+ }
370
+ //#endregion
371
+ //#region src/trace/capture-fetch.d.ts
372
+ interface CaptureFetchContext {
373
+ /** Logical run id stamped on every captured event. Required — without it
374
+ * the raw events can't be paired with their parent `Run`. */
375
+ runId: string;
376
+ /** Optional logical span id (enables span-level sink filtering). */
377
+ spanId?: string;
378
+ /** Resolved base URL (normalised, no trailing slash). Used for the event's
379
+ * `baseUrl` and for endpoint-path extraction. */
380
+ baseUrl: string;
381
+ /** Model id the caller intends to invoke. Stamped on every event. */
382
+ model: string;
383
+ /** Provider override. When omitted, `providerFromBaseUrl(baseUrl)`. */
384
+ provider?: string;
385
+ }
386
+ interface CaptureFetchOptions {
387
+ /** Override the capture-time redactor. Default `defaultProviderRedactor`. */
388
+ redactor?: ProviderRedactor;
389
+ /** Cap on captured response-body bytes; beyond it the body is truncated and
390
+ * `body_truncated` is added to `redactedFields`. Default 2 MiB. */
391
+ responseBodyByteCap?: number;
392
+ /** When true, a sink-write failure propagates to the caller. Default false
393
+ * — capture is best-effort so a sink failure never kills the LLM call. */
394
+ failClosed?: boolean;
395
+ /**
396
+ * Invoked with the token usage parsed off each successful response (JSON or
397
+ * SSE), keyed by the captured context. Lets a caller fold usage → cost without
398
+ * re-cloning the response themselves. Not called when the response carries no
399
+ * usage. Best-effort: a throw here is swallowed (it never kills the LLM call)
400
+ * unless `failClosed` is set.
401
+ */
402
+ onUsage?: (usage: ExtractedUsage, ctx: CaptureFetchContext) => void;
403
+ /** How repeated SSE usage events should be reconciled. Default `cumulative`. */
404
+ sseUsageMode?: SseUsageMode;
405
+ }
406
+ declare function captureFetchToRawSink(fetch: typeof globalThis.fetch, sink: RawProviderSink, ctx: CaptureFetchContext, opts?: CaptureFetchOptions): typeof globalThis.fetch;
407
+ //#endregion
407
408
  //#region src/trace-analyst/store-tool-spans.d.ts
408
409
  /** Missing tool spans cannot distinguish a tool-free run from broken capture. */
409
410
  declare class ToolTraceMissingError extends CaptureIntegrityError {
@@ -415,5 +416,5 @@ declare class ToolTraceMissingError extends CaptureIntegrityError {
415
416
  */
416
417
  declare function toolSpansToTraceAnalysisStore(spans: readonly ToolSpan[] | null | undefined, options?: ToolSpansToTraceAnalysisStoreOptions): TraceAnalysisStore;
417
418
  //#endregion
418
- export { scoreTraceInsightReadiness as A, OtlpSpan as B, describeTraceInsightScope as C, isTraceInsightAttribution as D, isTraceInsightArtifact as E, analyzeTraces as F, ExtractUsageFromSseOptions as G, CaptureFetchContext as H, OtlpFlatLine as I, extractUsage as J, ExtractedUsage as K, OTEL_AGENT_EVAL_SCOPE as L, AnalyzeTracesInput as M, AnalyzeTracesOptions as N, isTraceInsightEvidence as O, AnalyzeTracesResult as P, OtlpExport as R, defaultTraceInsightPanel as S, inferDomainKeywords as T, CaptureFetchOptions as U, exportRunAsOtlp as V, captureFetchToRawSink as W, extractUsageFromSse as X, extractUsageFromResponse as Y, TraceInsightReadiness as _, SpanRecordsToTraceAnalysisStoreOptions as a, buildTraceInsightContext as b, spanRecordsToTraceAnalysisStore as c, TraceInsightEvidence as d, TraceInsightFinding as f, TraceInsightQuestion as g, TraceInsightQualityGate as h, OtlpFileTraceStoreOptions as i, tokenizeDomainWords as j, planTraceInsightQuestions as k, TraceInsightAttribution as l, TraceInsightPromptInput as m, toolSpansToTraceAnalysisStore as n, ToolSpansToTraceAnalysisStoreOptions as o, TraceInsightPanelRole as p, SseUsageMode as q, OtlpFileTraceStore as r, otlpTextToTraceAnalysisStore as s, ToolTraceMissingError as t, TraceInsightContext as u, TraceInsightSuite as v, domainEvidencePattern as w, buildTraceInsightPrompt as x, TraceInsightTask as y, OtlpResourceSpans as z };
419
- //# sourceMappingURL=store-tool-spans-BNzls3OD.d.ts.map
419
+ export { TraceInsightQualityGate as A, inferDomainKeywords as B, exportRunAsOtlp as C, TraceInsightFinding as D, TraceInsightEvidence as E, buildTraceInsightContext as F, scoreTraceInsightReadiness as G, isTraceInsightAttribution as H, buildTraceInsightPrompt as I, AnalyzeTracesOptions as J, tokenizeDomainWords as K, defaultTraceInsightPanel as L, TraceInsightReadiness as M, TraceInsightSuite as N, TraceInsightPanelRole as O, TraceInsightTask as P, describeTraceInsightScope as R, OtlpSpan as S, TraceInsightContext as T, isTraceInsightEvidence as U, isTraceInsightArtifact as V, planTraceInsightQuestions as W, analyzeTraces as X, AnalyzeTracesResult as Y, spanRecordsToTraceAnalysisStore as _, captureFetchToRawSink as a, OtlpExport as b, ExtractedUsage as c, extractUsageFromResponse as d, extractUsageFromSse as f, otlpTextToTraceAnalysisStore as g, SpanRecordsToTraceAnalysisStoreOptions as h, CaptureFetchOptions as i, TraceInsightQuestion as j, TraceInsightPromptInput as k, SseUsageMode as l, OtlpFileTraceStoreOptions as m, toolSpansToTraceAnalysisStore as n, OtlpFlatLine as o, OtlpFileTraceStore as p, AnalyzeTracesInput as q, CaptureFetchContext as r, ExtractUsageFromSseOptions as s, ToolTraceMissingError as t, extractUsage as u, ToolSpansToTraceAnalysisStoreOptions as v, TraceInsightAttribution as w, OtlpResourceSpans as x, OTEL_AGENT_EVAL_SCOPE as y, domainEvidencePattern as z };
420
+ //# sourceMappingURL=store-tool-spans-DdCSPa7C.d.ts.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"store-tool-spans-DdCSPa7C.d.ts","names":[],"sources":["../src/trace-analyst/analyst.ts","../src/trace-analyst/insights.ts","../src/trace/otel.ts","../src/trace-analyst/store-otlp-buffered.ts","../src/trace-analyst/store-otlp.ts","../src/trace/extract-usage.ts","../src/trace/otlp-flat.ts","../src/trace/capture-fetch.ts","../src/trace-analyst/store-tool-spans.ts"],"mappings":";;;;;;;;;;;UAciB;EACf;EACA;EACA;EACA;;KAGU,sBAAsB;UAEjB;;EAEf,iBAAiB;;EAEjB,QAAQ;;EAER;;EAEA,YAAY;EACZ,SAAS,QAAQ;EACjB;EACA;EACA,aAAa;EACb;EACA,yBAAyB;EACzB,4BAA4B;EAC5B,eAAe,SAAS;EACxB,OAAO;EACP,OAAO,iBAAiB,SAAS;EACjC,SAAS;;;;;;;;iBASW,cACpB,OAAO,oBACP,SAAS,uBACR,QAAQ;;;UCnDM;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA,OAAO;;UAGQ;EACf;EACA;EACA;EACA;EACA;;EAEA,cAAc;;;UAIC;EACf;IACE;IACA;IACA;IACA;IACA;IACA;IACA;;EAEF,WAAW,gGAET;EAEF,OAAO;IAAQ;IAAiB;IAAgB,UAAU;;;KAGhD,0BAA0B;EAE9B;EAAoB;EAA0B;;EAE9C;EACA;EACA,OAAO;EACP,WAAW;;EAGX;EACA;EACA;EACA,OAAO;EACP,WAAW;;EAGX;EACA;EACA,OAAO;EACP,MAAM;EACN;EACA,mBAAmB;;;iBAaX,uBAAuB,iBAAiB,SAAS;iBAoBjD,uBAAuB,iBAAiB,SAAS;iBAyCjD,0BAA0B,iBAAiB,SAAS;UA8CnD;EACf;EACA;EACA;;UAGe;EACf;EACA;EACA;;UAGe;EACf,OAAO;EACP,WAAW;EACX,QAAQ;EACR,SAAS;EACT;;UAGe;EACf,OAAO;EACP;EACA;EACA,WAAW;EACX,OAAO;EACP,UAAU;EACV,OAAO;EACP,QAAQ;;UAGO;EACf;EACA;EACA;EACA;EACA;;UAGe;EACf;EACA;EACA,OAAO;;iBA0BO,oBAAoB;iBAMpB,oBAAoB,OAAO;iBAuB3B,sBAAsB,qBAAqB;iBAS3C,0BAA0B,OAAO;iBAsBjC,0BAA0B,OAAO,0BAA0B;iBAsD3D,yBAAyB,OAAO,0BAA0B;iBAa1D,2BAA2B,SAAS,sBAAsB;iBA+E1D,4BAA4B;iBA6B5B,wBAAwB,OAAO;;;cCvdlC;EAA0B;EAAoC;;UAE1D;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA,YAAY;IACV;IACA;MAAS;MAAsB;MAAmB;MAAsB;;;EAE1E,SAAS;IAAQ;IAAsB;IAAc,aAAa;;EAClE;IAAW;IAAc;;;UAGV;EACf;IAAY,YAAY;;EACxB,YAAY;IAAQ,cAAc;IAAuB,OAAO;;;UAGjD;EACf,eAAe;;;iBAIK,gBACpB,OAAO,YACP,eACA,gBAAe,4CACd,QAAQ;;;UC0CM;;EAEf,eAAe;;EAEf;;EAEA;;EAEA;;EAEA;;uBAuBoB,8BAA8B;WACzC,iBAAiB;mBACT;mBACA;mBACA;mBACA;UACT;UACA;UACA;UACA;EAER,YAAY,MAAM;EA+BZ,SAAS,kBAAkB,UAAU,4BAA4B;EAOjE,SACJ;IAAQ;IAAkB;KAC1B,UAAU,4BACT;EAeG,YACJ,UAAU,qBACV,UAAU,4BACT,QAAQ;EAoDL,YACJ;IAAQ,UAAU;IAAqB;IAAe;KACtD,UAAU,4BACT,QAAQ;EAmBL,YACJ,UAAU,qBACV,UAAU,4BACT;EAMG,UACJ;IAAQ;IAAkB;KAC1B,UAAU,4BACT,QAAQ;EAuCL,UACJ;IAAQ;IAAkB;IAA6B;KACvD,UAAU,4BACT,QAAQ;EAsFL,YACJ;IAAQ;IAAkB;IAAuB;KACjD,UAAU,4BACT,QAAQ;EA4CL,WACJ;IAAQ;IAAkB;IAAiB;IAAuB;KAClE,UAAU,4BACT,QAAQ;;;EA6CL,cAAc,UAAU,4BAA4B;UAI5C;qBAmBK,WAAW,SAAS,cAAc,QAAQ;qBAE1C;YAET,cAAc,KAAK,0BAA0B;UAIzC;UAmBA;UAwHA;UA+DN;UAkBA;UAwDA;UAmBM;;;;UC5xBC,kCAAkC;;EAEjD;;;;;;;;EAQA;;;KAIU,yCAAyC;cAKxC,2BAA2B;mBACrB;mBACA;EAEjB,YAAY,MAAM;;;YAYF,WAAW,SAAS,cAAc,QAAQ;YAkBhD;;;;;;;iBAkCI,gCACd,kBAAkB,oBAClB,UAAS,yCACR;;iBA6Ba,6BACd,cACA,OAAM,uCACL;;;KChJS,iBAAiB;KACjB;UAEK;;EAEf,OAAO;;;;;;;iBAuDO,aAAa,gBAAgB;;;;;;iBAkD7B,oBACd,cACA,UAAS,6BACR;;;;;;;iBA4DmB,yBACpB,UAAU,UACV,aAAa,6BACZ,QAAQ;;;KC5LC;UAEK;EACf;EACA;EACA;EACA;EACA;EACA;EACA;EACA;IACE,MAAM;IACN;;EAEF;IAAY,YAAY;;EACxB,YAAY;EACZ,SAAS;IAAQ;IAAc;IAAuB,aAAa;;;;;UCEpD;;;EAGf;;EAEA;;;EAGA;;EAEA;;EAEA;;UAGe;;EAEf,WAAW;;;EAGX;;;EAGA;;;;;;;;EAQA,WAAW,OAAO,gBAAgB,KAAK;;EAEvC,eAAe;;iBAkDD,sBACd,cAAc,WAAW,OACzB,MAAM,iBACN,KAAK,qBACL,OAAM,6BACE,WAAW;;;;cCrGR,8BAA8B;EACzC;;;;;;iBASc,8BACd,gBAAgB,+BAChB,UAAS,uCACR"}
@@ -1 +1 @@
1
- {"version":3,"file":"index.d.ts","names":[],"sources":["../../src/storyboard/code-edit.ts","../../src/storyboard/index.ts"],"mappings":";;;iBAuBiB;;EAEf;EACA;EACA;;EAEA;;;EAGA;EACA;EACA;EACA;EACA;;;wBA6Dc,iBAAiB,MAAM,OAAO;;wBAwB9B,iBAAiB,gBAAgB,SAAS;;;wBAS1C,uBACd,YAAY,YACZ,gBAAgB,SACf;EAAQ,OAAO;EAAO,MAAM;;;;;wBAgBf,kBAAkB,MAAM;iBAiBvB;;EAEf;;EAEA;;EAEA;;;;;;;;;wBAUc,wBACd,iBAAiB,kBAAkB,YACnC,OAAM;;;;YCjKI;;;YAmBA;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;;iBAEW;EACf,MAAM;;EAEN;;EAEA;;EAEA,QAAQ;;EAER;;EAEA;EACA;EACA;;iBAGe;;EAEf;;;;;;;;;;wBAiTc,uBACd,gBAAgB,QAChB,OAAM,gBACL;YA8CS;iBAiBK;EACf,WAAW;EACX;EACA;;EAEA,QAAQ;EACR;EACA;;iBAGe;EACf;EACA;EACA,QAAQ;;iBAGO;;EAEf;;EAEA;;;;;;;wBAkCc,kBACd,iBAAiB,iBACjB,OAAM,iBACL;;wBA4Ja,yBAAyB,YAAY;iBAwBpC;;EAEf;;;;;;;;;wBAUc,qBAAqB,YAAY,YAAY,OAAM"}
1
+ {"version":3,"file":"index.d.ts","names":[],"sources":["../../src/storyboard/code-edit.ts","../../src/storyboard/index.ts"],"mappings":";;;iBAuBiB;;EAEf;EACA;EACA;;EAEA;;;EAGA;EACA;EACA;EACA;EACA;;;wBA6Dc,iBAAiB,MAAM,OAAO;;wBAwB9B,iBAAiB,gBAAgB,SAAS;;;wBAS1C,uBACd,YAAY,YACZ,gBAAgB,SACf;EAAQ,OAAO;EAAO,MAAM;;;;;wBAgBf,kBAAkB,MAAM;iBAiBvB;;EAEf;;EAEA;;EAEA;;;;;;;;;wBAUc,wBACd,iBAAiB,kBAAkB,YACnC,OAAM;;;;YChKI;;;YAmBA;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;;iBAEW;EACf,MAAM;;EAEN;;EAEA;;EAEA,QAAQ;;EAER;;EAEA;EACA;EACA;;iBAGe;;EAEf;;;;;;;;;;wBAiTc,uBACd,gBAAgB,QAChB,OAAM,gBACL;YA8CS;iBAiBK;EACf,WAAW;EACX;EACA;;EAEA,QAAQ;EACR;EACA;;iBAGe;EACf;EACA;EACA,QAAQ;;iBAGO;;EAEf;;EAEA;;;;;;;wBAkCc,kBACd,iBAAiB,iBACjB,OAAM,iBACL;;wBAoJa,yBAAyB,YAAY;iBAwBpC;;EAEf;;;;;;;;;wBAUc,qBAAqB,YAAY,YAAY,OAAM"}
@@ -1,3 +1,4 @@
1
+ import { t as escapeHtml } from "../reporting-html-Kvlc8oMt.js";
1
2
  //#region src/storyboard/code-edit.ts
2
3
  const EDIT_TOOLS = /(edit|write|patch|apply|str_replace|create.*file|save|insert|update.*file|multi_edit)/i;
3
4
  function languageOf(path) {
@@ -227,6 +228,25 @@ function renderCodeAnimationHtml(source, opts = {}) {
227
228
  }
228
229
  //#endregion
229
230
  //#region src/storyboard/index.ts
231
+ /**
232
+ * Storyboard — compile a raw agent trace into a small set of meaningful scenes,
233
+ * then render them as a shareable, watchable replay.
234
+ *
235
+ * Span[] → SemanticEvent[] → Storyboard{scenes} → { markdown | html }
236
+ *
237
+ * The trace is the source of truth; the storyboard is the IR; markdown/html
238
+ * (and, in a consumer package, Remotion/MP4) are compiled targets. Everything
239
+ * here is PURE — same spans in, same bytes out (no clock/random) — so a run's
240
+ * replay is deterministic and diffable.
241
+ *
242
+ * Crucially, every scene carries a modality-typed VISUAL extracted from the
243
+ * span — a browser/computer screenshot, a code diff, a terminal command +
244
+ * output, file content, an API request/response, or reasoning prose — so no
245
+ * matter what the agent did, the replay SHOWS it, not just narrates it. The
246
+ * reducer is rules-based; an LLM pass can refine titles/summaries later, but
247
+ * the structure + visuals do not depend on one. Renderers that need a browser
248
+ * (Remotion, React) live in consumers; this module only emits strings.
249
+ */
230
250
  function str(v) {
231
251
  return typeof v === "string" && v.length > 0 ? v : void 0;
232
252
  }
@@ -648,9 +668,6 @@ function mmss(ms) {
648
668
  function truncate(s, max) {
649
669
  return s.length <= max ? s : `${s.slice(0, Math.max(0, max - 1))}…`;
650
670
  }
651
- function escapeHtml(s) {
652
- return s.replace(/&/g, "&amp;").replace(/</g, "&lt;").replace(/>/g, "&gt;").replace(/"/g, "&quot;");
653
- }
654
671
  /** A short modality tag for the markdown shot list. */
655
672
  function visualMarkdown(v) {
656
673
  switch (v.type) {