promptfoo 0.121.1 → 0.121.3

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 (316) hide show
  1. package/README.md +2 -0
  2. package/dist/src/{accounts-xrUGFA6n.js → accounts-B2XmGjty.js} +5 -5
  3. package/dist/src/{accounts-Bx-x3bmW.cjs → accounts-BPyfpSeU.cjs} +5 -5
  4. package/dist/src/{accounts-CMqkzrVf.js → accounts-CFLK3mnD.js} +6 -6
  5. package/dist/src/{accounts-BgNJDBE6.js → accounts-Xatc0RYb.js} +5 -5
  6. package/dist/src/{agentic-utils-BKIN5PKu.js → agentic-utils-36epdqwB.js} +3 -3
  7. package/dist/src/{cometapi-DkXrKi5z.js → agentic-utils-D8yXo5Lm.js} +4 -61
  8. package/dist/src/{cometapi-vY6aDZgo.cjs → agentic-utils-DAVsChuB.cjs} +24 -62
  9. package/dist/src/agentic-utils-DIYAAYE7.js +153 -0
  10. package/dist/src/{agents-C-dDThPK.js → agents-BBVJCIYr.js} +226 -13
  11. package/dist/src/{agents-CErsqg5U.cjs → agents-BBWxKSM0.cjs} +7 -7
  12. package/dist/src/{agents-Dy2YpZpa.js → agents-Bqgfdokm.js} +227 -14
  13. package/dist/src/{agents-B0f4HICh.cjs → agents-CAYbM7qD.cjs} +226 -13
  14. package/dist/src/{agents-CVIn-Utx.js → agents-CLQ-P15P.js} +7 -7
  15. package/dist/src/{agents-DeH4Gu94.js → agents-CgBniSlI.js} +8 -8
  16. package/dist/src/{agents-CXknwsFX.js → agents-DSSTV4bv.js} +226 -13
  17. package/dist/src/{agents-aF4-T121.js → agents-wg3ohknq.js} +7 -7
  18. package/dist/src/{aimlapi-tg0Gkcvr.cjs → aimlapi-Bv8Fmc-b.cjs} +14 -14
  19. package/dist/src/{aimlapi-BNfTBexL.js → aimlapi-BwGC1TtS.js} +13 -13
  20. package/dist/src/{aimlapi-BAGZDo5G.js → aimlapi-DaC3qZ-o.js} +14 -14
  21. package/dist/src/{aimlapi-DHRKlBEA.js → aimlapi-MgSLdvy7.js} +13 -13
  22. package/dist/src/app/assets/index-B6l9CVVb.js +439 -0
  23. package/dist/src/app/assets/index-DyZ0Ep37.css +1 -0
  24. package/dist/src/app/assets/sync-CStkzc6u.js +4 -0
  25. package/dist/src/app/assets/vendor-charts-BnDWwBlI.js +36 -0
  26. package/dist/src/app/assets/vendor-markdown-Bz7N-ca6.js +29 -0
  27. package/dist/src/app/index.html +4 -4
  28. package/dist/src/{audio-tf_NBjlC.js → audio-Bn44pQxv.js} +4 -4
  29. package/dist/src/{audio-CHQ4r-RV.js → audio-DDA5WHdx.js} +4 -4
  30. package/dist/src/{audio-BWeaWovU.cjs → audio-DVFjQ67_.cjs} +4 -4
  31. package/dist/src/{audio-BRODU0UK.js → audio-DjU9GswO.js} +5 -5
  32. package/dist/src/{base-DBtwl2FR.cjs → base-BboXIF_0.cjs} +3 -3
  33. package/dist/src/{base-B4QJRyFS.js → base-CKjwebIH.js} +3 -3
  34. package/dist/src/{base-B0tcrnq_.js → base-CqzQ4K8j.js} +3 -3
  35. package/dist/src/{base-fEDN28WM.js → base-Cz2ZC_iA.js} +3 -3
  36. package/dist/src/{blobs-BAU-dXan.js → blobs-B1JriOyi.js} +3 -3
  37. package/dist/src/{blobs-qTYm-1PY.js → blobs-BUWmKWzo.js} +3 -3
  38. package/dist/src/{blobs-DvS-O6be.cjs → blobs-C6j0bvFz.cjs} +3 -3
  39. package/dist/src/{blobs-Bpg5rH6i.js → blobs-DXTl6J3H.js} +3 -3
  40. package/dist/src/{cache-COish3-W.cjs → cache-C5yFZ4gC.cjs} +75 -58
  41. package/dist/src/{cache-8XhNqPKW.js → cache-CaT5tPgo.js} +75 -58
  42. package/dist/src/cache-CyCanoMu.js +6 -0
  43. package/dist/src/{cache-CG0SlR1d.js → cache-DSqR6ezl.js} +75 -58
  44. package/dist/src/cache-Df_QFDNu.cjs +5 -0
  45. package/dist/src/{cache-D3eqDYGU.js → cache-HP0NP4k3.js} +75 -58
  46. package/dist/src/{chat-DHMH-N64.js → chat-B-52XYI1.js} +12 -12
  47. package/dist/src/{chat-BKm79wib.js → chat-B0iaWhoh.js} +16 -14
  48. package/dist/src/{chat-DxysjBvt.js → chat-BE0qTA8e.js} +13 -13
  49. package/dist/src/{chat-CRWNNq73.js → chat-BEwdgGEg.js} +16 -14
  50. package/dist/src/{chat-2K608PeQ.cjs → chat-BtIKkLKx.cjs} +13 -13
  51. package/dist/src/{chat-DaqekjFr.cjs → chat-CM8qWR3_.cjs} +17 -15
  52. package/dist/src/{chat-CM_kyI8B.js → chat-DK1U-eZ-.js} +12 -12
  53. package/dist/src/{chat-CznLWr_D.js → chat-pxmiVpWe.js} +16 -14
  54. package/dist/src/{chatkit-65VXf5SR.js → chatkit-BYGQlHlV.js} +4 -4
  55. package/dist/src/{chatkit-DKyPi1Gs.cjs → chatkit-Cx174XI3.cjs} +4 -4
  56. package/dist/src/{chatkit-BxFvW8KY.js → chatkit-_8eJqKcD.js} +4 -4
  57. package/dist/src/{chatkit-Be-Q-a9F.js → chatkit-a2D6mY6s.js} +4 -4
  58. package/dist/src/{claude-agent-sdk-CJH22shf.cjs → claude-agent-sdk-8ddRp1L2.cjs} +35 -17
  59. package/dist/src/{claude-agent-sdk-Dy5lT-Tx.js → claude-agent-sdk-Bq5EArsX.js} +33 -15
  60. package/dist/src/{claude-agent-sdk-BLTu0WBO.js → claude-agent-sdk-CMjh4LFH.js} +33 -15
  61. package/dist/src/{claude-agent-sdk-D6_k9FKA.js → claude-agent-sdk-HgbFioFw.js} +33 -15
  62. package/dist/src/cloud-DE3t1-ZI.js +4 -0
  63. package/dist/src/{cloud-Bc9526yV.js → cloud-z8KZpUoa.js} +3 -3
  64. package/dist/src/{cloudflare-ai-CWWJCRim.js → cloudflare-ai-BGyXlpXJ.js} +13 -13
  65. package/dist/src/{cloudflare-ai-C9r2sRhw.js → cloudflare-ai-Bbp26N0L.js} +13 -13
  66. package/dist/src/{cloudflare-ai-ClWSdor4.cjs → cloudflare-ai-C62x6MQG.cjs} +14 -14
  67. package/dist/src/{cloudflare-ai-ICsOuD-z.js → cloudflare-ai-DdKP9TKT.js} +14 -14
  68. package/dist/src/{cloudflare-gateway-D6xFc5pa.js → cloudflare-gateway-BwAaUgeW.js} +14 -14
  69. package/dist/src/{cloudflare-gateway-D6O7AlYb.js → cloudflare-gateway-D-e9i1Sn.js} +15 -15
  70. package/dist/src/{cloudflare-gateway-pXGHxJ47.js → cloudflare-gateway-DXhtXDRb.js} +15 -163
  71. package/dist/src/{cloudflare-gateway-C2_-KG5o.cjs → cloudflare-gateway-Dx36ftqF.cjs} +15 -15
  72. package/dist/src/{codex-sdk-DUwKWezN.js → codex-sdk-BQEw16R_.js} +180 -11
  73. package/dist/src/{codex-sdk-C6UMlxwV.js → codex-sdk-C_07GuVS.js} +180 -11
  74. package/dist/src/{codex-sdk-GGAw0qbD.js → codex-sdk-DE5G18dx.js} +180 -11
  75. package/dist/src/{codex-sdk-fAO0c3yA.cjs → codex-sdk-ZLKfDjqP.cjs} +181 -12
  76. package/dist/src/cometapi-BDyV-NNm.js +62 -0
  77. package/dist/src/cometapi-C3hOlM7-.cjs +62 -0
  78. package/dist/src/{cometapi-Bbjp5V4x.js → cometapi-hhL4TAh3.js} +14 -14
  79. package/dist/src/{cometapi-BasUi7-_.js → cometapi-sp7sJpBD.js} +15 -15
  80. package/dist/src/{completion-C_P3ypkJ.js → completion-BCimtq-h.js} +6 -6
  81. package/dist/src/{completion-6Mx_iXxK.js → completion-DCjv7RZ3.js} +6 -6
  82. package/dist/src/{completion-CDOouNzq.cjs → completion-DlXUhj5c.cjs} +6 -6
  83. package/dist/src/{completion-C5rtR_9P.js → completion-DoYy49ti.js} +6 -6
  84. package/dist/src/{createHash-CfZSc0b4.cjs → createHash-BYwImsYv.cjs} +2 -2
  85. package/dist/src/{docker-BwsKwxFs.cjs → docker-Cqj2-QVi.cjs} +14 -14
  86. package/dist/src/{docker-CZnqU1XV.js → docker-CxCkwMzc.js} +13 -13
  87. package/dist/src/{docker-DzxyDPIj.js → docker-DpguQj-w.js} +14 -14
  88. package/dist/src/{docker-5KcG-_86.js → docker-FeBni2dw.js} +13 -13
  89. package/dist/src/{esm-C03C-mv3.js → esm-7UIl0pPM.js} +2 -2
  90. package/dist/src/{esm-Cd1AjG1D.js → esm-CKWP3u_P.js} +3 -3
  91. package/dist/src/{esm-CnNt7sI4.cjs → esm-CipptfDu.cjs} +2 -2
  92. package/dist/src/{esm-CaIwzWR5.js → esm-SUNIX1x3.js} +3 -3
  93. package/dist/src/eval-7aEqoMs3.js +15 -0
  94. package/dist/src/{eval-DmFyWU7i.js → eval-BTqTn7lb.js} +10 -10
  95. package/dist/src/{evalResult-CDQiuUuf.js → evalResult-BkIhRdTe.js} +7 -7
  96. package/dist/src/evalResult-CYNHkk5A.js +12 -0
  97. package/dist/src/evalResult-CuvJeNiM.js +10 -0
  98. package/dist/src/{evalResult-CTG2AHOS.js → evalResult-DUDShQrm.js} +7 -7
  99. package/dist/src/{evalResult-Dap2CekP.cjs → evalResult-DpARzUCb.cjs} +7 -7
  100. package/dist/src/evalResult-tGdilrWt.cjs +10 -0
  101. package/dist/src/evaluator-BBUqRhz1.js +36 -0
  102. package/dist/src/{evaluator-DPFRbFIL.js → evaluator-BcvOGaam.js} +833 -79
  103. package/dist/src/{extractor-YMU_Gvt8.js → extractor-C8XwivI9.js} +6 -6
  104. package/dist/src/{extractor-CFG6bcWJ.js → extractor-CAZ2G3Kh.js} +6 -6
  105. package/dist/src/{extractor-DX36oYEv.cjs → extractor-DG3sSfXE.cjs} +6 -6
  106. package/dist/src/{extractor-M67RUtg6.js → extractor-D_wd8jxt.js} +6 -6
  107. package/dist/src/{fetch-4M3YRaqL.js → fetch-BiYv2BZc.js} +3 -3
  108. package/dist/src/{fetch-BxUk8odA.cjs → fetch-BnR9wSnm.cjs} +3 -3
  109. package/dist/src/{fetch-60Gzydls.js → fetch-CVAtKnI3.js} +3 -3
  110. package/dist/src/{fetch-BMv0O527.js → fetch-DoVRJZhJ.js} +4 -4
  111. package/dist/src/fetch-UWU706qb.js +5 -0
  112. package/dist/src/{genaiTracer-DN4dQywX.cjs → genaiTracer-BfxrvSUb.cjs} +2 -2
  113. package/dist/src/{graders-DOXycdlG.cjs → graders-BElhu9ZY.cjs} +126 -55
  114. package/dist/src/{graders-R9rYUM0d.js → graders-BXAJ0sbS.js} +120 -55
  115. package/dist/src/graders-BxfEguVY.js +32 -0
  116. package/dist/src/graders-CzVMbEnv.js +34 -0
  117. package/dist/src/{graders-CpdqD9PI.js → graders-DG7mhg-b.js} +120 -55
  118. package/dist/src/graders-DjCXfj0l.cjs +32 -0
  119. package/dist/src/{graders-CHO8EPM4.js → graders-RjHF8VfG.js} +120 -55
  120. package/dist/src/graders-kHzIWOKu.js +32 -0
  121. package/dist/src/{image-DTedmQPg.cjs → image--F58eEIn.cjs} +6 -6
  122. package/dist/src/{image-DJEvKveK.js → image-6WQXK8m8.js} +4 -4
  123. package/dist/src/{image-pAX56tPG.js → image-B8b6f36E.js} +6 -6
  124. package/dist/src/{image-BmEZqVmk.js → image-CoxZp9PZ.js} +6 -6
  125. package/dist/src/{image-gvmivTEe.js → image-DO0RYnjH.js} +5 -5
  126. package/dist/src/{image-CBBVXWuT.js → image-PoF6DN3x.js} +6 -6
  127. package/dist/src/{image-CDLQOcqT.cjs → image-fza3zuKs.cjs} +4 -4
  128. package/dist/src/{image-tL5hIOFh.js → image-xNbw5ph2.js} +4 -4
  129. package/dist/src/index.cjs +863 -110
  130. package/dist/src/index.d.cts +833 -60
  131. package/dist/src/index.d.ts +833 -60
  132. package/dist/src/index.js +860 -108
  133. package/dist/src/{interactiveCheck-BgLZUIt3.js → interactiveCheck-BnMYOjMu.js} +2 -2
  134. package/dist/src/{knowledgeBase-CoU-UQBg.js → knowledgeBase-Bi7CmDbx.js} +7 -7
  135. package/dist/src/{knowledgeBase-CLJybhnF.js → knowledgeBase-Ce3ofVan.js} +8 -8
  136. package/dist/src/{knowledgeBase-DjWPVqSb.js → knowledgeBase-DFRXPZl_.js} +7 -7
  137. package/dist/src/{knowledgeBase-wkxuRFhA.cjs → knowledgeBase-DqrLX8fy.cjs} +7 -7
  138. package/dist/src/{litellm-B9Hysuri.js → litellm-Bo2gQXpo.js} +16 -15
  139. package/dist/src/{litellm-ePxtr9F1.js → litellm-CKiAxnoM.js} +15 -14
  140. package/dist/src/{litellm-NYpQ8RQu.cjs → litellm-CnHI69aj.cjs} +16 -15
  141. package/dist/src/{litellm-CTfa0hqi.js → litellm-Tc294Jhj.js} +15 -14
  142. package/dist/src/{logger-KkObSCzq.js → logger-BcJBzSSA.js} +10 -14
  143. package/dist/src/{logger-DLcq4dWf.js → logger-BnkjG2jt.js} +10 -14
  144. package/dist/src/{logger-Cp1GPUjj.cjs → logger-D5iKBpu_.cjs} +27 -13
  145. package/dist/src/{logger-CT3IKMKA.js → logger-DO8_zM18.js} +10 -14
  146. package/dist/src/{luma-ray-BW9IRGIc.js → luma-ray-0ehMPt5N.js} +10 -10
  147. package/dist/src/{luma-ray-BE2mOt6N.js → luma-ray-C9q8rdQe.js} +9 -9
  148. package/dist/src/{luma-ray-Cm1KZBhs.js → luma-ray-DP0QA9qn.js} +9 -9
  149. package/dist/src/{luma-ray-B0GGNRc1.cjs → luma-ray-m9Ku2meV.cjs} +9 -9
  150. package/dist/src/main.js +69 -71
  151. package/dist/src/{messages-1x9atZmP.js → messages-DJNo37Ko.js} +14 -9
  152. package/dist/src/{messages-BLbWdsyt.js → messages-Dy9QecMs.js} +14 -9
  153. package/dist/src/{messages-1JrJs91T.cjs → messages-HJsyEh4o.cjs} +15 -10
  154. package/dist/src/{messages-D8EA0oDc.js → messages-biC_ex-p.js} +14 -9
  155. package/dist/src/{modelslab-C1OLRmVX.js → modelslab-B5J-ZM5c.js} +9 -9
  156. package/dist/src/{modelslab-CqXBy3U8.js → modelslab-BI458moT.js} +10 -10
  157. package/dist/src/{modelslab-X5-4LroM.js → modelslab-BTOT8FUO.js} +9 -9
  158. package/dist/src/{modelslab-DcOSFwKh.cjs → modelslab-IQbNg-r7.cjs} +9 -9
  159. package/dist/src/{nova-reel-DihqLeol.js → nova-reel-BZ9y-Y5s.js} +9 -9
  160. package/dist/src/{nova-reel-D9xfaMBs.cjs → nova-reel-CE5etkv9.cjs} +9 -9
  161. package/dist/src/{nova-reel-D2ZkOSyr.js → nova-reel-DEeQlnOJ.js} +10 -10
  162. package/dist/src/{nova-reel-BgS1ZWuK.js → nova-reel-Xw1SXLpg.js} +9 -9
  163. package/dist/src/{nova-sonic-Q3BOJeig.js → nova-sonic-DWswpN1E.js} +7 -7
  164. package/dist/src/{nova-sonic-DezhVUYT.js → nova-sonic-DXTLpi-r.js} +6 -6
  165. package/dist/src/{nova-sonic-DVu3mMIy.cjs → nova-sonic-N0yCm0vb.cjs} +6 -6
  166. package/dist/src/{nova-sonic-P-CdUMlV.js → nova-sonic-Ogqf-csn.js} +6 -6
  167. package/dist/src/{openai-DhbB7eWK.js → openai-BMcwgD5C.js} +2 -2
  168. package/dist/src/{openai-j-sE2O7r.js → openai-BcB5KlTk.js} +2 -2
  169. package/dist/src/{openai-Cuif0GEt.cjs → openai-CoxGAQwn.cjs} +2 -2
  170. package/dist/src/{openai-DElQ-fPX.js → openai-D6wITiVn.js} +2 -2
  171. package/dist/src/{openclaw-Bv1DINsX.js → openclaw-0Sv7AK3O.js} +172 -109
  172. package/dist/src/{openclaw-DAfWQn-o.cjs → openclaw-CXxbKgDH.cjs} +174 -110
  173. package/dist/src/{openclaw-BiSZPL7J.js → openclaw-D1FSCps-.js} +172 -109
  174. package/dist/src/{openclaw-D1D_ej1z.js → openclaw-D2ENvu7a.js} +173 -110
  175. package/dist/src/{opencode-sdk-D95s6SnR.js → opencode-sdk-C71Z0ehR.js} +13 -13
  176. package/dist/src/{opencode-sdk-DxUPkLT7.js → opencode-sdk-CHCs7dEb.js} +12 -12
  177. package/dist/src/{opencode-sdk-C7m-wRfI.js → opencode-sdk-DDxj4QqH.js} +12 -12
  178. package/dist/src/{opencode-sdk-CfaLN8PY.cjs → opencode-sdk-WWJhnbKr.cjs} +16 -16
  179. package/dist/src/{otlpReceiver-g3ByGaXs.js → otlpReceiver-C9KlUtxh.js} +6 -6
  180. package/dist/src/{otlpReceiver--AIRW_S4.js → otlpReceiver-CZL48YfC.js} +6 -6
  181. package/dist/src/{otlpReceiver-Bn5wGB1v.js → otlpReceiver-CavGAA6k.js} +6 -6
  182. package/dist/src/{otlpReceiver-Diec4cln.cjs → otlpReceiver-DHKqJlsz.cjs} +6 -6
  183. package/dist/src/{providerRegistry-B0RUOLI_.js → providerRegistry-B9lh-_tx.js} +2 -2
  184. package/dist/src/{providerRegistry-Civky8Ar.cjs → providerRegistry-BTDgfV5h.cjs} +2 -2
  185. package/dist/src/{providerRegistry-CD8MEar9.js → providerRegistry-BkzVH5Ba.js} +2 -2
  186. package/dist/src/{providerRegistry-DM8rZYol.js → providerRegistry-CUWki5mQ.js} +2 -2
  187. package/dist/src/providers-BSLEaIQG.js +32 -0
  188. package/dist/src/{providers-CFu-TZl-.cjs → providers-CScd1wN6.cjs} +733 -464
  189. package/dist/src/{providers-CFLy1_ji.js → providers-Ch6Mr0gn.js} +795 -526
  190. package/dist/src/{providers-BKRJTjBz.js → providers-Cn73d5sr.js} +795 -526
  191. package/dist/src/providers-D-FnDg8k.cjs +31 -0
  192. package/dist/src/providers-DEYiFVAo.js +30 -0
  193. package/dist/src/{providers-B3HvufyI.js → providers-DvddrgxL.js} +795 -526
  194. package/dist/src/providers-sS2WI8YD.js +30 -0
  195. package/dist/src/{pythonUtils-D6fwaDSg.js → pythonUtils-Bzwbgpbg.js} +3 -3
  196. package/dist/src/{pythonUtils-D5nxkQ0P.js → pythonUtils-Cpo0Ez1p.js} +3 -3
  197. package/dist/src/{pythonUtils-CTU3Y3lw.cjs → pythonUtils-dAVigVK-.cjs} +3 -3
  198. package/dist/src/{pythonUtils-C3py6GC1.js → pythonUtils-wIqk7zAf.js} +3 -3
  199. package/dist/src/{quiverai-CI6gYJVI.js → quiverai-BeofbLVc.js} +4 -4
  200. package/dist/src/{quiverai-MHSxbmmZ.js → quiverai-CCQn73lq.js} +5 -5
  201. package/dist/src/{quiverai-CLkWkyZc.cjs → quiverai-CcUhPIBg.cjs} +4 -4
  202. package/dist/src/{quiverai-C2jVwbH1.js → quiverai-DVSEqJiq.js} +4 -4
  203. package/dist/src/{render-Drod8m7K.js → render-BHl6QVq9.js} +3 -3
  204. package/dist/src/{responses-CGw0DCzh.js → responses-BKP_WYis.js} +16 -12
  205. package/dist/src/{responses-BKqJmhhc.js → responses-CQb1Tj69.js} +16 -12
  206. package/dist/src/{responses-jxdehPkC.js → responses-CgNyTPsY.js} +16 -12
  207. package/dist/src/{responses-tD4Bd4dc.cjs → responses-mo0KQDbu.cjs} +16 -12
  208. package/dist/src/rubyUtils-B1HXG4ej.cjs +4 -0
  209. package/dist/src/{rubyUtils-DhCAlxZr.cjs → rubyUtils-CGeUtCfW.cjs} +3 -3
  210. package/dist/src/{rubyUtils-Boc4HZzX.js → rubyUtils-CiVfln3g.js} +3 -3
  211. package/dist/src/{rubyUtils-BcuGX77l.js → rubyUtils-DECSbsfY.js} +3 -3
  212. package/dist/src/{rubyUtils-BUVePouc.js → rubyUtils-PgU-gHmx.js} +3 -3
  213. package/dist/src/rubyUtils-Rt6pKA96.js +5 -0
  214. package/dist/src/{sagemaker-BK4Zb993.js → sagemaker-CVv8W7so.js} +17 -17
  215. package/dist/src/{sagemaker-D2Q1c-sD.js → sagemaker-CqeASYE5.js} +17 -17
  216. package/dist/src/{sagemaker-BfiWTmvn.js → sagemaker-MUbD5V3v.js} +18 -18
  217. package/dist/src/{sagemaker-CcQHM1jV.cjs → sagemaker-jiw1wQa-.cjs} +17 -17
  218. package/dist/src/{scanner-J8CA3LsV.js → scanner-DVDeUz1r.js} +10 -10
  219. package/dist/src/server/index.js +864 -112
  220. package/dist/src/server-B0Xh1Gx-.js +7 -0
  221. package/dist/src/{server-B0PPuDw-.cjs → server-BtoCXeXI.cjs} +4 -4
  222. package/dist/src/{server-BC7XJFgr.js → server-CP9qKM40.js} +4 -4
  223. package/dist/src/{server-OAs3nBRT.js → server-Cns05F1j.js} +5 -5
  224. package/dist/src/server-DJTKu9IR.cjs +5 -0
  225. package/dist/src/{server-DbFphssR.js → server-DZ9MtCn0.js} +6 -6
  226. package/dist/src/{signal-BOTbd53Z.js → signal-C3ZTsUgi.js} +3 -3
  227. package/dist/src/{slack-DXMKtA-f.js → slack-2sdpGzbt.js} +2 -2
  228. package/dist/src/{slack-BmVAVGaK.cjs → slack-94iG3T0s.cjs} +2 -2
  229. package/dist/src/{slack-DCUPTzS2.js → slack-BR0HtO3K.js} +2 -2
  230. package/dist/src/{slack-DOdy_kyv.js → slack-DCEV-vWP.js} +2 -2
  231. package/dist/src/store-C5u6MgC8.js +6 -0
  232. package/dist/src/{store-BSc-TF2w.cjs → store-CLyU7AtI.cjs} +17 -5
  233. package/dist/src/store-CNHk-De4.cjs +5 -0
  234. package/dist/src/{store-DQLEjuEO.js → store-Cj258DgL.js} +17 -5
  235. package/dist/src/{store-D1tv90v3.js → store-P8OKm19S.js} +17 -5
  236. package/dist/src/{store-Ub2vaGJ1.js → store-VB0GP46K.js} +17 -5
  237. package/dist/src/{tables-xKANLRBD.js → tables-BEIFz2tM.js} +3 -3
  238. package/dist/src/{tables-C7K-XKWp.cjs → tables-BdZQEpRz.cjs} +3 -3
  239. package/dist/src/{tables-D36WTqKX.js → tables-DmzvLbeZ.js} +3 -3
  240. package/dist/src/{tables-5EvT_Bwn.js → tables-kC7R5kiK.js} +3 -3
  241. package/dist/src/{telemetry-C2YDkUQH.js → telemetry-BnH5VJAU.js} +4 -4
  242. package/dist/src/{telemetry-C15ziL8u.js → telemetry-BugWqKiu.js} +4 -4
  243. package/dist/src/{telemetry-DMb2Mpfm.js → telemetry-DPXLd7UE.js} +4 -4
  244. package/dist/src/telemetry-Yig0Tino.js +7 -0
  245. package/dist/src/telemetry-p8Pwqm1i.cjs +5 -0
  246. package/dist/src/{telemetry-CbrnxHp_.cjs → telemetry-re627Lre.cjs} +4 -4
  247. package/dist/src/{transcription-CL78qbOU.cjs → transcription-BvtsrzRG.cjs} +13 -13
  248. package/dist/src/{transcription-DAtxHhAM.js → transcription-CaMivnjG.js} +13 -13
  249. package/dist/src/{transcription-QHh3AH6Z.js → transcription-DOMMTu01.js} +14 -14
  250. package/dist/src/{transcription-LNZTNUUL.js → transcription-Hb3VnC4M.js} +13 -13
  251. package/dist/src/{transform-DOcQeLld.cjs → transform-0BwoBsvO.cjs} +19 -5
  252. package/dist/src/{transform-DGxXocjk.js → transform-B2-jIv68.js} +8 -6
  253. package/dist/src/{transform-DECvGmzp.js → transform-BqPkNPYm.js} +4 -4
  254. package/dist/src/{transform-aa6tmVpZ.js → transform-BzK09Q_9.js} +4 -4
  255. package/dist/src/transform-ChNIpHz7.js +6 -0
  256. package/dist/src/{transform-Cgi24fJ7.js → transform-DrleutM3.js} +8 -6
  257. package/dist/src/{transform-DGLazrMm.js → transform-DyDAwEpE.js} +8 -6
  258. package/dist/src/transform-PtQ6rAE3.cjs +5 -0
  259. package/dist/src/{transform-CzK1Q0zl.cjs → transform-ZrG2dvlo.cjs} +4 -4
  260. package/dist/src/{transform-DilY9wbS.js → transform-ljLYHEPh.js} +4 -4
  261. package/dist/src/{transformersAvailability-CEVM2GNQ.js → transformersAvailability-BGkzavwb.js} +1 -1
  262. package/dist/src/{transformersAvailability-CwayUSlh.cjs → transformersAvailability-DKoRtQLy.cjs} +1 -1
  263. package/dist/src/{types-CH3Ge2sE.js → types-CIhFeUC4.js} +45 -11
  264. package/dist/src/{types-CN_TZ2GJ.js → types-Cd3ygw8W.js} +45 -11
  265. package/dist/src/{types-LJ0r3wbR.cjs → types-D8cGDZbL.cjs} +46 -12
  266. package/dist/src/{types-CLKiCBW3.js → types-q8GXGF65.js} +45 -11
  267. package/dist/src/{util-CchiqXh_.cjs → util--9u9UVCt.cjs} +3 -3
  268. package/dist/src/{util-5cB-L7U3.js → util-BLvy9qfE.js} +7 -11
  269. package/dist/src/{util-YT5HPZaS.js → util-Bm3E9jpK.js} +7 -11
  270. package/dist/src/{util-6-GqIvzS.js → util-BtoGs5Cb.js} +18 -4
  271. package/dist/src/{util-Db0a0AFH.cjs → util-CFj4YKIn.cjs} +18 -4
  272. package/dist/src/{util-Dlz_Wvgm.js → util-CMMkIxfU.js} +7 -11
  273. package/dist/src/{util-Betm42rL.js → util-CgDCK4KI.js} +18 -4
  274. package/dist/src/{util-Yz-1aEhW.cjs → util-CuLo2pMR.cjs} +7 -11
  275. package/dist/src/{util-C-PPYSMq.js → util-DM2rTn_6.js} +18 -4
  276. package/dist/src/{util-B7T3SiBS.js → util-DMFeUvLz.js} +3 -3
  277. package/dist/src/{util-ZZH-3QZz.js → util-DbVG-yZU.js} +3 -3
  278. package/dist/src/{util-DaWTWKBK.js → util-vNmDL5DT.js} +3 -3
  279. package/dist/src/{utils-XiOAgly5.js → utils-CFxO9KGo.js} +2 -2
  280. package/dist/src/{utils-f2-Moju7.js → utils-DEuL4VNB.js} +2 -2
  281. package/dist/src/{utils-Cz9qXqII.cjs → utils-DKw8mrgr.cjs} +3 -3
  282. package/dist/src/{utils-dLokC-eR.js → utils-DOjD4dTC.js} +2 -2
  283. package/dist/tsconfig.tsbuildinfo +1 -1
  284. package/package.json +38 -38
  285. package/dist/src/app/assets/index-BFCZg7hQ.js +0 -439
  286. package/dist/src/app/assets/index-NCn4eVBv.css +0 -1
  287. package/dist/src/app/assets/sync-9qqYcY-B.js +0 -4
  288. package/dist/src/app/assets/vendor-charts-CCl15Imd.js +0 -36
  289. package/dist/src/app/assets/vendor-markdown-0tekx3KX.js +0 -29
  290. package/dist/src/cache-Bbn1Nyrd.cjs +0 -5
  291. package/dist/src/cache-BwsMSda7.js +0 -6
  292. package/dist/src/cloud-DmE0EwsY.js +0 -4
  293. package/dist/src/eval-17JizQIv.js +0 -15
  294. package/dist/src/evalResult-Cqj8pldJ.js +0 -12
  295. package/dist/src/evalResult-DvcJAWJU.cjs +0 -10
  296. package/dist/src/evalResult-Hftn-S_i.js +0 -10
  297. package/dist/src/evaluator-B2CFNt-P.js +0 -36
  298. package/dist/src/fetch-KV5kNASw.js +0 -5
  299. package/dist/src/graders-Bu0H9nXi.js +0 -32
  300. package/dist/src/graders-Cfhkvx-e.js +0 -34
  301. package/dist/src/graders-DClJVpGP.cjs +0 -32
  302. package/dist/src/graders-DcnJsrMO.js +0 -32
  303. package/dist/src/providers-C1rOSHiR.js +0 -32
  304. package/dist/src/providers-CxmDwEFf.cjs +0 -31
  305. package/dist/src/providers-Dodakqr0.js +0 -30
  306. package/dist/src/providers-GIQ2TcsA.js +0 -30
  307. package/dist/src/rubyUtils-BUHu6PhO.js +0 -5
  308. package/dist/src/rubyUtils-CP42kMvq.cjs +0 -4
  309. package/dist/src/server-B1vi21hA.js +0 -7
  310. package/dist/src/server-Cm9Kai_h.cjs +0 -5
  311. package/dist/src/store-BNmZ1KAz.cjs +0 -5
  312. package/dist/src/store-BltJg2cd.js +0 -6
  313. package/dist/src/telemetry-5BCRNBbe.cjs +0 -5
  314. package/dist/src/telemetry-D4W5hboe.js +0 -7
  315. package/dist/src/transform-DTGDnAzW.js +0 -6
  316. package/dist/src/transform-m3qNw4KP.cjs +0 -5
package/dist/src/index.d.ts CHANGED
@@ -89,7 +89,12 @@ declare const ProviderEnvOverridesSchema: z.ZodObject<{
89
89
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
90
90
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
91
91
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
92
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
93
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
94
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
92
95
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
96
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
97
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
93
98
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
94
99
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
95
100
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -238,6 +243,13 @@ interface ChatMessage {
238
243
  role: 'system' | 'user' | 'assistant' | 'tool' | 'function';
239
244
  content: string;
240
245
  }
246
+ interface SkillCallEntry {
247
+ name: string;
248
+ input?: unknown;
249
+ path?: string;
250
+ source?: 'heuristic' | 'tool';
251
+ is_error?: boolean;
252
+ }
241
253
  type ProviderTypeMap = Partial<Record<ProviderType, string | ProviderOptions | ApiProvider>>;
242
254
  interface AtomicTestCase$1 {
243
255
  description?: string;
@@ -575,7 +587,12 @@ declare const ProvidersSchema: z.ZodUnion<readonly [z.ZodString, z.ZodCustom<Cal
575
587
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
576
588
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
577
589
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
590
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
591
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
592
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
578
593
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
594
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
595
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
579
596
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
580
597
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
581
598
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -702,7 +719,12 @@ declare const ProvidersSchema: z.ZodUnion<readonly [z.ZodString, z.ZodCustom<Cal
702
719
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
703
720
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
704
721
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
722
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
723
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
724
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
705
725
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
726
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
727
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
706
728
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
707
729
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
708
730
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -1647,6 +1669,12 @@ declare const BaseAssertionTypesSchema: z.ZodEnum<{
1647
1669
  "similar:euclidean": "similar:euclidean";
1648
1670
  "starts-with": "starts-with";
1649
1671
  "tool-call-f1": "tool-call-f1";
1672
+ "skill-used": "skill-used";
1673
+ "trajectory:goal-success": "trajectory:goal-success";
1674
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
1675
+ "trajectory:step-count": "trajectory:step-count";
1676
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
1677
+ "trajectory:tool-used": "trajectory:tool-used";
1650
1678
  "trace-error-spans": "trace-error-spans";
1651
1679
  "trace-span-count": "trace-span-count";
1652
1680
  "trace-span-duration": "trace-span-duration";
@@ -1715,13 +1743,19 @@ declare const NotPrefixedAssertionTypesSchema: z.ZodPipe<z.ZodEnum<{
1715
1743
  "similar:euclidean": "similar:euclidean";
1716
1744
  "starts-with": "starts-with";
1717
1745
  "tool-call-f1": "tool-call-f1";
1746
+ "skill-used": "skill-used";
1747
+ "trajectory:goal-success": "trajectory:goal-success";
1748
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
1749
+ "trajectory:step-count": "trajectory:step-count";
1750
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
1751
+ "trajectory:tool-used": "trajectory:tool-used";
1718
1752
  "trace-error-spans": "trace-error-spans";
1719
1753
  "trace-span-count": "trace-span-count";
1720
1754
  "trace-span-duration": "trace-span-duration";
1721
1755
  "search-rubric": "search-rubric";
1722
1756
  webhook: "webhook";
1723
1757
  "word-count": "word-count";
1724
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>;
1758
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>;
1725
1759
  declare const AssertionTypeSchema: z.ZodUnion<readonly [z.ZodEnum<{
1726
1760
  "answer-relevance": "answer-relevance";
1727
1761
  bleu: "bleu";
@@ -1776,6 +1810,12 @@ declare const AssertionTypeSchema: z.ZodUnion<readonly [z.ZodEnum<{
1776
1810
  "similar:euclidean": "similar:euclidean";
1777
1811
  "starts-with": "starts-with";
1778
1812
  "tool-call-f1": "tool-call-f1";
1813
+ "skill-used": "skill-used";
1814
+ "trajectory:goal-success": "trajectory:goal-success";
1815
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
1816
+ "trajectory:step-count": "trajectory:step-count";
1817
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
1818
+ "trajectory:tool-used": "trajectory:tool-used";
1779
1819
  "trace-error-spans": "trace-error-spans";
1780
1820
  "trace-span-count": "trace-span-count";
1781
1821
  "trace-span-duration": "trace-span-duration";
@@ -1836,13 +1876,19 @@ declare const AssertionTypeSchema: z.ZodUnion<readonly [z.ZodEnum<{
1836
1876
  "similar:euclidean": "similar:euclidean";
1837
1877
  "starts-with": "starts-with";
1838
1878
  "tool-call-f1": "tool-call-f1";
1879
+ "skill-used": "skill-used";
1880
+ "trajectory:goal-success": "trajectory:goal-success";
1881
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
1882
+ "trajectory:step-count": "trajectory:step-count";
1883
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
1884
+ "trajectory:tool-used": "trajectory:tool-used";
1839
1885
  "trace-error-spans": "trace-error-spans";
1840
1886
  "trace-span-count": "trace-span-count";
1841
1887
  "trace-span-duration": "trace-span-duration";
1842
1888
  "search-rubric": "search-rubric";
1843
1889
  webhook: "webhook";
1844
1890
  "word-count": "word-count";
1845
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
1891
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
1846
1892
  "select-best": "select-best";
1847
1893
  human: "human";
1848
1894
  "max-score": "max-score";
@@ -1905,6 +1951,12 @@ declare const AssertionSetSchema: z.ZodObject<{
1905
1951
  "similar:euclidean": "similar:euclidean";
1906
1952
  "starts-with": "starts-with";
1907
1953
  "tool-call-f1": "tool-call-f1";
1954
+ "skill-used": "skill-used";
1955
+ "trajectory:goal-success": "trajectory:goal-success";
1956
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
1957
+ "trajectory:step-count": "trajectory:step-count";
1958
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
1959
+ "trajectory:tool-used": "trajectory:tool-used";
1908
1960
  "trace-error-spans": "trace-error-spans";
1909
1961
  "trace-span-count": "trace-span-count";
1910
1962
  "trace-span-duration": "trace-span-duration";
@@ -1965,13 +2017,19 @@ declare const AssertionSetSchema: z.ZodObject<{
1965
2017
  "similar:euclidean": "similar:euclidean";
1966
2018
  "starts-with": "starts-with";
1967
2019
  "tool-call-f1": "tool-call-f1";
2020
+ "skill-used": "skill-used";
2021
+ "trajectory:goal-success": "trajectory:goal-success";
2022
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
2023
+ "trajectory:step-count": "trajectory:step-count";
2024
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
2025
+ "trajectory:tool-used": "trajectory:tool-used";
1968
2026
  "trace-error-spans": "trace-error-spans";
1969
2027
  "trace-span-count": "trace-span-count";
1970
2028
  "trace-span-duration": "trace-span-duration";
1971
2029
  "search-rubric": "search-rubric";
1972
2030
  webhook: "webhook";
1973
2031
  "word-count": "word-count";
1974
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
2032
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
1975
2033
  "select-best": "select-best";
1976
2034
  human: "human";
1977
2035
  "max-score": "max-score";
@@ -2053,6 +2111,12 @@ declare const AssertionSchema: z.ZodObject<{
2053
2111
  "similar:euclidean": "similar:euclidean";
2054
2112
  "starts-with": "starts-with";
2055
2113
  "tool-call-f1": "tool-call-f1";
2114
+ "skill-used": "skill-used";
2115
+ "trajectory:goal-success": "trajectory:goal-success";
2116
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
2117
+ "trajectory:step-count": "trajectory:step-count";
2118
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
2119
+ "trajectory:tool-used": "trajectory:tool-used";
2056
2120
  "trace-error-spans": "trace-error-spans";
2057
2121
  "trace-span-count": "trace-span-count";
2058
2122
  "trace-span-duration": "trace-span-duration";
@@ -2113,13 +2177,19 @@ declare const AssertionSchema: z.ZodObject<{
2113
2177
  "similar:euclidean": "similar:euclidean";
2114
2178
  "starts-with": "starts-with";
2115
2179
  "tool-call-f1": "tool-call-f1";
2180
+ "skill-used": "skill-used";
2181
+ "trajectory:goal-success": "trajectory:goal-success";
2182
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
2183
+ "trajectory:step-count": "trajectory:step-count";
2184
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
2185
+ "trajectory:tool-used": "trajectory:tool-used";
2116
2186
  "trace-error-spans": "trace-error-spans";
2117
2187
  "trace-span-count": "trace-span-count";
2118
2188
  "trace-span-duration": "trace-span-duration";
2119
2189
  "search-rubric": "search-rubric";
2120
2190
  webhook: "webhook";
2121
2191
  "word-count": "word-count";
2122
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
2192
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
2123
2193
  "select-best": "select-best";
2124
2194
  human: "human";
2125
2195
  "max-score": "max-score";
@@ -2202,6 +2272,12 @@ declare const AssertionOrSetSchema: z.ZodUnion<readonly [z.ZodObject<{
2202
2272
  "similar:euclidean": "similar:euclidean";
2203
2273
  "starts-with": "starts-with";
2204
2274
  "tool-call-f1": "tool-call-f1";
2275
+ "skill-used": "skill-used";
2276
+ "trajectory:goal-success": "trajectory:goal-success";
2277
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
2278
+ "trajectory:step-count": "trajectory:step-count";
2279
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
2280
+ "trajectory:tool-used": "trajectory:tool-used";
2205
2281
  "trace-error-spans": "trace-error-spans";
2206
2282
  "trace-span-count": "trace-span-count";
2207
2283
  "trace-span-duration": "trace-span-duration";
@@ -2262,13 +2338,19 @@ declare const AssertionOrSetSchema: z.ZodUnion<readonly [z.ZodObject<{
2262
2338
  "similar:euclidean": "similar:euclidean";
2263
2339
  "starts-with": "starts-with";
2264
2340
  "tool-call-f1": "tool-call-f1";
2341
+ "skill-used": "skill-used";
2342
+ "trajectory:goal-success": "trajectory:goal-success";
2343
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
2344
+ "trajectory:step-count": "trajectory:step-count";
2345
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
2346
+ "trajectory:tool-used": "trajectory:tool-used";
2265
2347
  "trace-error-spans": "trace-error-spans";
2266
2348
  "trace-span-count": "trace-span-count";
2267
2349
  "trace-span-duration": "trace-span-duration";
2268
2350
  "search-rubric": "search-rubric";
2269
2351
  webhook: "webhook";
2270
2352
  "word-count": "word-count";
2271
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
2353
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
2272
2354
  "select-best": "select-best";
2273
2355
  human: "human";
2274
2356
  "max-score": "max-score";
@@ -2348,6 +2430,12 @@ declare const AssertionOrSetSchema: z.ZodUnion<readonly [z.ZodObject<{
2348
2430
  "similar:euclidean": "similar:euclidean";
2349
2431
  "starts-with": "starts-with";
2350
2432
  "tool-call-f1": "tool-call-f1";
2433
+ "skill-used": "skill-used";
2434
+ "trajectory:goal-success": "trajectory:goal-success";
2435
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
2436
+ "trajectory:step-count": "trajectory:step-count";
2437
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
2438
+ "trajectory:tool-used": "trajectory:tool-used";
2351
2439
  "trace-error-spans": "trace-error-spans";
2352
2440
  "trace-span-count": "trace-span-count";
2353
2441
  "trace-span-duration": "trace-span-duration";
@@ -2408,13 +2496,19 @@ declare const AssertionOrSetSchema: z.ZodUnion<readonly [z.ZodObject<{
2408
2496
  "similar:euclidean": "similar:euclidean";
2409
2497
  "starts-with": "starts-with";
2410
2498
  "tool-call-f1": "tool-call-f1";
2499
+ "skill-used": "skill-used";
2500
+ "trajectory:goal-success": "trajectory:goal-success";
2501
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
2502
+ "trajectory:step-count": "trajectory:step-count";
2503
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
2504
+ "trajectory:tool-used": "trajectory:tool-used";
2411
2505
  "trace-error-spans": "trace-error-spans";
2412
2506
  "trace-span-count": "trace-span-count";
2413
2507
  "trace-span-duration": "trace-span-duration";
2414
2508
  "search-rubric": "search-rubric";
2415
2509
  webhook: "webhook";
2416
2510
  "word-count": "word-count";
2417
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
2511
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
2418
2512
  "select-best": "select-best";
2419
2513
  human: "human";
2420
2514
  "max-score": "max-score";
@@ -2622,7 +2716,12 @@ declare const TestCaseSchema: z.ZodObject<{
2622
2716
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
2623
2717
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
2624
2718
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
2719
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
2720
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
2721
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
2625
2722
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
2723
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
2724
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
2626
2725
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
2627
2726
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
2628
2727
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -2741,6 +2840,12 @@ declare const TestCaseSchema: z.ZodObject<{
2741
2840
  "similar:euclidean": "similar:euclidean";
2742
2841
  "starts-with": "starts-with";
2743
2842
  "tool-call-f1": "tool-call-f1";
2843
+ "skill-used": "skill-used";
2844
+ "trajectory:goal-success": "trajectory:goal-success";
2845
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
2846
+ "trajectory:step-count": "trajectory:step-count";
2847
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
2848
+ "trajectory:tool-used": "trajectory:tool-used";
2744
2849
  "trace-error-spans": "trace-error-spans";
2745
2850
  "trace-span-count": "trace-span-count";
2746
2851
  "trace-span-duration": "trace-span-duration";
@@ -2801,13 +2906,19 @@ declare const TestCaseSchema: z.ZodObject<{
2801
2906
  "similar:euclidean": "similar:euclidean";
2802
2907
  "starts-with": "starts-with";
2803
2908
  "tool-call-f1": "tool-call-f1";
2909
+ "skill-used": "skill-used";
2910
+ "trajectory:goal-success": "trajectory:goal-success";
2911
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
2912
+ "trajectory:step-count": "trajectory:step-count";
2913
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
2914
+ "trajectory:tool-used": "trajectory:tool-used";
2804
2915
  "trace-error-spans": "trace-error-spans";
2805
2916
  "trace-span-count": "trace-span-count";
2806
2917
  "trace-span-duration": "trace-span-duration";
2807
2918
  "search-rubric": "search-rubric";
2808
2919
  webhook: "webhook";
2809
2920
  "word-count": "word-count";
2810
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
2921
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
2811
2922
  "select-best": "select-best";
2812
2923
  human: "human";
2813
2924
  "max-score": "max-score";
@@ -2887,6 +2998,12 @@ declare const TestCaseSchema: z.ZodObject<{
2887
2998
  "similar:euclidean": "similar:euclidean";
2888
2999
  "starts-with": "starts-with";
2889
3000
  "tool-call-f1": "tool-call-f1";
3001
+ "skill-used": "skill-used";
3002
+ "trajectory:goal-success": "trajectory:goal-success";
3003
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
3004
+ "trajectory:step-count": "trajectory:step-count";
3005
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
3006
+ "trajectory:tool-used": "trajectory:tool-used";
2890
3007
  "trace-error-spans": "trace-error-spans";
2891
3008
  "trace-span-count": "trace-span-count";
2892
3009
  "trace-span-duration": "trace-span-duration";
@@ -2947,13 +3064,19 @@ declare const TestCaseSchema: z.ZodObject<{
2947
3064
  "similar:euclidean": "similar:euclidean";
2948
3065
  "starts-with": "starts-with";
2949
3066
  "tool-call-f1": "tool-call-f1";
3067
+ "skill-used": "skill-used";
3068
+ "trajectory:goal-success": "trajectory:goal-success";
3069
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
3070
+ "trajectory:step-count": "trajectory:step-count";
3071
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
3072
+ "trajectory:tool-used": "trajectory:tool-used";
2950
3073
  "trace-error-spans": "trace-error-spans";
2951
3074
  "trace-span-count": "trace-span-count";
2952
3075
  "trace-span-duration": "trace-span-duration";
2953
3076
  "search-rubric": "search-rubric";
2954
3077
  webhook: "webhook";
2955
3078
  "word-count": "word-count";
2956
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
3079
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
2957
3080
  "select-best": "select-best";
2958
3081
  human: "human";
2959
3082
  "max-score": "max-score";
@@ -3167,7 +3290,12 @@ declare const TestCaseWithVarsFileSchema: z.ZodObject<{
3167
3290
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
3168
3291
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
3169
3292
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
3293
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
3294
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
3295
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
3170
3296
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
3297
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
3298
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
3171
3299
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
3172
3300
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
3173
3301
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -3286,6 +3414,12 @@ declare const TestCaseWithVarsFileSchema: z.ZodObject<{
3286
3414
  "similar:euclidean": "similar:euclidean";
3287
3415
  "starts-with": "starts-with";
3288
3416
  "tool-call-f1": "tool-call-f1";
3417
+ "skill-used": "skill-used";
3418
+ "trajectory:goal-success": "trajectory:goal-success";
3419
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
3420
+ "trajectory:step-count": "trajectory:step-count";
3421
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
3422
+ "trajectory:tool-used": "trajectory:tool-used";
3289
3423
  "trace-error-spans": "trace-error-spans";
3290
3424
  "trace-span-count": "trace-span-count";
3291
3425
  "trace-span-duration": "trace-span-duration";
@@ -3346,13 +3480,19 @@ declare const TestCaseWithVarsFileSchema: z.ZodObject<{
3346
3480
  "similar:euclidean": "similar:euclidean";
3347
3481
  "starts-with": "starts-with";
3348
3482
  "tool-call-f1": "tool-call-f1";
3483
+ "skill-used": "skill-used";
3484
+ "trajectory:goal-success": "trajectory:goal-success";
3485
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
3486
+ "trajectory:step-count": "trajectory:step-count";
3487
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
3488
+ "trajectory:tool-used": "trajectory:tool-used";
3349
3489
  "trace-error-spans": "trace-error-spans";
3350
3490
  "trace-span-count": "trace-span-count";
3351
3491
  "trace-span-duration": "trace-span-duration";
3352
3492
  "search-rubric": "search-rubric";
3353
3493
  webhook: "webhook";
3354
3494
  "word-count": "word-count";
3355
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
3495
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
3356
3496
  "select-best": "select-best";
3357
3497
  human: "human";
3358
3498
  "max-score": "max-score";
@@ -3432,6 +3572,12 @@ declare const TestCaseWithVarsFileSchema: z.ZodObject<{
3432
3572
  "similar:euclidean": "similar:euclidean";
3433
3573
  "starts-with": "starts-with";
3434
3574
  "tool-call-f1": "tool-call-f1";
3575
+ "skill-used": "skill-used";
3576
+ "trajectory:goal-success": "trajectory:goal-success";
3577
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
3578
+ "trajectory:step-count": "trajectory:step-count";
3579
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
3580
+ "trajectory:tool-used": "trajectory:tool-used";
3435
3581
  "trace-error-spans": "trace-error-spans";
3436
3582
  "trace-span-count": "trace-span-count";
3437
3583
  "trace-span-duration": "trace-span-duration";
@@ -3492,13 +3638,19 @@ declare const TestCaseWithVarsFileSchema: z.ZodObject<{
3492
3638
  "similar:euclidean": "similar:euclidean";
3493
3639
  "starts-with": "starts-with";
3494
3640
  "tool-call-f1": "tool-call-f1";
3641
+ "skill-used": "skill-used";
3642
+ "trajectory:goal-success": "trajectory:goal-success";
3643
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
3644
+ "trajectory:step-count": "trajectory:step-count";
3645
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
3646
+ "trajectory:tool-used": "trajectory:tool-used";
3495
3647
  "trace-error-spans": "trace-error-spans";
3496
3648
  "trace-span-count": "trace-span-count";
3497
3649
  "trace-span-duration": "trace-span-duration";
3498
3650
  "search-rubric": "search-rubric";
3499
3651
  webhook: "webhook";
3500
3652
  "word-count": "word-count";
3501
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
3653
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
3502
3654
  "select-best": "select-best";
3503
3655
  human: "human";
3504
3656
  "max-score": "max-score";
@@ -3711,7 +3863,12 @@ declare const TestCasesWithMetadataSchema: z.ZodObject<{
3711
3863
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
3712
3864
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
3713
3865
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
3866
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
3867
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
3868
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
3714
3869
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
3870
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
3871
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
3715
3872
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
3716
3873
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
3717
3874
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -3830,6 +3987,12 @@ declare const TestCasesWithMetadataSchema: z.ZodObject<{
3830
3987
  "similar:euclidean": "similar:euclidean";
3831
3988
  "starts-with": "starts-with";
3832
3989
  "tool-call-f1": "tool-call-f1";
3990
+ "skill-used": "skill-used";
3991
+ "trajectory:goal-success": "trajectory:goal-success";
3992
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
3993
+ "trajectory:step-count": "trajectory:step-count";
3994
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
3995
+ "trajectory:tool-used": "trajectory:tool-used";
3833
3996
  "trace-error-spans": "trace-error-spans";
3834
3997
  "trace-span-count": "trace-span-count";
3835
3998
  "trace-span-duration": "trace-span-duration";
@@ -3890,13 +4053,19 @@ declare const TestCasesWithMetadataSchema: z.ZodObject<{
3890
4053
  "similar:euclidean": "similar:euclidean";
3891
4054
  "starts-with": "starts-with";
3892
4055
  "tool-call-f1": "tool-call-f1";
4056
+ "skill-used": "skill-used";
4057
+ "trajectory:goal-success": "trajectory:goal-success";
4058
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
4059
+ "trajectory:step-count": "trajectory:step-count";
4060
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
4061
+ "trajectory:tool-used": "trajectory:tool-used";
3893
4062
  "trace-error-spans": "trace-error-spans";
3894
4063
  "trace-span-count": "trace-span-count";
3895
4064
  "trace-span-duration": "trace-span-duration";
3896
4065
  "search-rubric": "search-rubric";
3897
4066
  webhook: "webhook";
3898
4067
  "word-count": "word-count";
3899
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
4068
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
3900
4069
  "select-best": "select-best";
3901
4070
  human: "human";
3902
4071
  "max-score": "max-score";
@@ -3976,6 +4145,12 @@ declare const TestCasesWithMetadataSchema: z.ZodObject<{
3976
4145
  "similar:euclidean": "similar:euclidean";
3977
4146
  "starts-with": "starts-with";
3978
4147
  "tool-call-f1": "tool-call-f1";
4148
+ "skill-used": "skill-used";
4149
+ "trajectory:goal-success": "trajectory:goal-success";
4150
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
4151
+ "trajectory:step-count": "trajectory:step-count";
4152
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
4153
+ "trajectory:tool-used": "trajectory:tool-used";
3979
4154
  "trace-error-spans": "trace-error-spans";
3980
4155
  "trace-span-count": "trace-span-count";
3981
4156
  "trace-span-duration": "trace-span-duration";
@@ -4036,13 +4211,19 @@ declare const TestCasesWithMetadataSchema: z.ZodObject<{
4036
4211
  "similar:euclidean": "similar:euclidean";
4037
4212
  "starts-with": "starts-with";
4038
4213
  "tool-call-f1": "tool-call-f1";
4214
+ "skill-used": "skill-used";
4215
+ "trajectory:goal-success": "trajectory:goal-success";
4216
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
4217
+ "trajectory:step-count": "trajectory:step-count";
4218
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
4219
+ "trajectory:tool-used": "trajectory:tool-used";
4039
4220
  "trace-error-spans": "trace-error-spans";
4040
4221
  "trace-span-count": "trace-span-count";
4041
4222
  "trace-span-duration": "trace-span-duration";
4042
4223
  "search-rubric": "search-rubric";
4043
4224
  webhook: "webhook";
4044
4225
  "word-count": "word-count";
4045
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
4226
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
4046
4227
  "select-best": "select-best";
4047
4228
  human: "human";
4048
4229
  "max-score": "max-score";
@@ -4313,7 +4494,12 @@ declare const ScenarioSchema: z.ZodObject<{
4313
4494
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
4314
4495
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
4315
4496
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
4497
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
4498
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
4499
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
4316
4500
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
4501
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
4502
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
4317
4503
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
4318
4504
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
4319
4505
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -4432,6 +4618,12 @@ declare const ScenarioSchema: z.ZodObject<{
4432
4618
  "similar:euclidean": "similar:euclidean";
4433
4619
  "starts-with": "starts-with";
4434
4620
  "tool-call-f1": "tool-call-f1";
4621
+ "skill-used": "skill-used";
4622
+ "trajectory:goal-success": "trajectory:goal-success";
4623
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
4624
+ "trajectory:step-count": "trajectory:step-count";
4625
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
4626
+ "trajectory:tool-used": "trajectory:tool-used";
4435
4627
  "trace-error-spans": "trace-error-spans";
4436
4628
  "trace-span-count": "trace-span-count";
4437
4629
  "trace-span-duration": "trace-span-duration";
@@ -4492,13 +4684,19 @@ declare const ScenarioSchema: z.ZodObject<{
4492
4684
  "similar:euclidean": "similar:euclidean";
4493
4685
  "starts-with": "starts-with";
4494
4686
  "tool-call-f1": "tool-call-f1";
4687
+ "skill-used": "skill-used";
4688
+ "trajectory:goal-success": "trajectory:goal-success";
4689
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
4690
+ "trajectory:step-count": "trajectory:step-count";
4691
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
4692
+ "trajectory:tool-used": "trajectory:tool-used";
4495
4693
  "trace-error-spans": "trace-error-spans";
4496
4694
  "trace-span-count": "trace-span-count";
4497
4695
  "trace-span-duration": "trace-span-duration";
4498
4696
  "search-rubric": "search-rubric";
4499
4697
  webhook: "webhook";
4500
4698
  "word-count": "word-count";
4501
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
4699
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
4502
4700
  "select-best": "select-best";
4503
4701
  human: "human";
4504
4702
  "max-score": "max-score";
@@ -4578,6 +4776,12 @@ declare const ScenarioSchema: z.ZodObject<{
4578
4776
  "similar:euclidean": "similar:euclidean";
4579
4777
  "starts-with": "starts-with";
4580
4778
  "tool-call-f1": "tool-call-f1";
4779
+ "skill-used": "skill-used";
4780
+ "trajectory:goal-success": "trajectory:goal-success";
4781
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
4782
+ "trajectory:step-count": "trajectory:step-count";
4783
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
4784
+ "trajectory:tool-used": "trajectory:tool-used";
4581
4785
  "trace-error-spans": "trace-error-spans";
4582
4786
  "trace-span-count": "trace-span-count";
4583
4787
  "trace-span-duration": "trace-span-duration";
@@ -4638,13 +4842,19 @@ declare const ScenarioSchema: z.ZodObject<{
4638
4842
  "similar:euclidean": "similar:euclidean";
4639
4843
  "starts-with": "starts-with";
4640
4844
  "tool-call-f1": "tool-call-f1";
4845
+ "skill-used": "skill-used";
4846
+ "trajectory:goal-success": "trajectory:goal-success";
4847
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
4848
+ "trajectory:step-count": "trajectory:step-count";
4849
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
4850
+ "trajectory:tool-used": "trajectory:tool-used";
4641
4851
  "trace-error-spans": "trace-error-spans";
4642
4852
  "trace-span-count": "trace-span-count";
4643
4853
  "trace-span-duration": "trace-span-duration";
4644
4854
  "search-rubric": "search-rubric";
4645
4855
  webhook: "webhook";
4646
4856
  "word-count": "word-count";
4647
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
4857
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
4648
4858
  "select-best": "select-best";
4649
4859
  human: "human";
4650
4860
  "max-score": "max-score";
@@ -4853,7 +5063,12 @@ declare const ScenarioSchema: z.ZodObject<{
4853
5063
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
4854
5064
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
4855
5065
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
5066
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
5067
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
5068
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
4856
5069
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
5070
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
5071
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
4857
5072
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
4858
5073
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
4859
5074
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -4972,6 +5187,12 @@ declare const ScenarioSchema: z.ZodObject<{
4972
5187
  "similar:euclidean": "similar:euclidean";
4973
5188
  "starts-with": "starts-with";
4974
5189
  "tool-call-f1": "tool-call-f1";
5190
+ "skill-used": "skill-used";
5191
+ "trajectory:goal-success": "trajectory:goal-success";
5192
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
5193
+ "trajectory:step-count": "trajectory:step-count";
5194
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
5195
+ "trajectory:tool-used": "trajectory:tool-used";
4975
5196
  "trace-error-spans": "trace-error-spans";
4976
5197
  "trace-span-count": "trace-span-count";
4977
5198
  "trace-span-duration": "trace-span-duration";
@@ -5032,13 +5253,19 @@ declare const ScenarioSchema: z.ZodObject<{
5032
5253
  "similar:euclidean": "similar:euclidean";
5033
5254
  "starts-with": "starts-with";
5034
5255
  "tool-call-f1": "tool-call-f1";
5256
+ "skill-used": "skill-used";
5257
+ "trajectory:goal-success": "trajectory:goal-success";
5258
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
5259
+ "trajectory:step-count": "trajectory:step-count";
5260
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
5261
+ "trajectory:tool-used": "trajectory:tool-used";
5035
5262
  "trace-error-spans": "trace-error-spans";
5036
5263
  "trace-span-count": "trace-span-count";
5037
5264
  "trace-span-duration": "trace-span-duration";
5038
5265
  "search-rubric": "search-rubric";
5039
5266
  webhook: "webhook";
5040
5267
  "word-count": "word-count";
5041
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
5268
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
5042
5269
  "select-best": "select-best";
5043
5270
  human: "human";
5044
5271
  "max-score": "max-score";
@@ -5118,6 +5345,12 @@ declare const ScenarioSchema: z.ZodObject<{
5118
5345
  "similar:euclidean": "similar:euclidean";
5119
5346
  "starts-with": "starts-with";
5120
5347
  "tool-call-f1": "tool-call-f1";
5348
+ "skill-used": "skill-used";
5349
+ "trajectory:goal-success": "trajectory:goal-success";
5350
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
5351
+ "trajectory:step-count": "trajectory:step-count";
5352
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
5353
+ "trajectory:tool-used": "trajectory:tool-used";
5121
5354
  "trace-error-spans": "trace-error-spans";
5122
5355
  "trace-span-count": "trace-span-count";
5123
5356
  "trace-span-duration": "trace-span-duration";
@@ -5178,13 +5411,19 @@ declare const ScenarioSchema: z.ZodObject<{
5178
5411
  "similar:euclidean": "similar:euclidean";
5179
5412
  "starts-with": "starts-with";
5180
5413
  "tool-call-f1": "tool-call-f1";
5414
+ "skill-used": "skill-used";
5415
+ "trajectory:goal-success": "trajectory:goal-success";
5416
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
5417
+ "trajectory:step-count": "trajectory:step-count";
5418
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
5419
+ "trajectory:tool-used": "trajectory:tool-used";
5181
5420
  "trace-error-spans": "trace-error-spans";
5182
5421
  "trace-span-count": "trace-span-count";
5183
5422
  "trace-span-duration": "trace-span-duration";
5184
5423
  "search-rubric": "search-rubric";
5185
5424
  webhook: "webhook";
5186
5425
  "word-count": "word-count";
5187
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
5426
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
5188
5427
  "select-best": "select-best";
5189
5428
  human: "human";
5190
5429
  "max-score": "max-score";
@@ -5394,7 +5633,12 @@ declare const AtomicTestCaseSchema: z.ZodObject<{
5394
5633
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
5395
5634
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
5396
5635
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
5636
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
5637
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
5638
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
5397
5639
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
5640
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
5641
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
5398
5642
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
5399
5643
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
5400
5644
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -5513,6 +5757,12 @@ declare const AtomicTestCaseSchema: z.ZodObject<{
5513
5757
  "similar:euclidean": "similar:euclidean";
5514
5758
  "starts-with": "starts-with";
5515
5759
  "tool-call-f1": "tool-call-f1";
5760
+ "skill-used": "skill-used";
5761
+ "trajectory:goal-success": "trajectory:goal-success";
5762
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
5763
+ "trajectory:step-count": "trajectory:step-count";
5764
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
5765
+ "trajectory:tool-used": "trajectory:tool-used";
5516
5766
  "trace-error-spans": "trace-error-spans";
5517
5767
  "trace-span-count": "trace-span-count";
5518
5768
  "trace-span-duration": "trace-span-duration";
@@ -5573,13 +5823,19 @@ declare const AtomicTestCaseSchema: z.ZodObject<{
5573
5823
  "similar:euclidean": "similar:euclidean";
5574
5824
  "starts-with": "starts-with";
5575
5825
  "tool-call-f1": "tool-call-f1";
5826
+ "skill-used": "skill-used";
5827
+ "trajectory:goal-success": "trajectory:goal-success";
5828
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
5829
+ "trajectory:step-count": "trajectory:step-count";
5830
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
5831
+ "trajectory:tool-used": "trajectory:tool-used";
5576
5832
  "trace-error-spans": "trace-error-spans";
5577
5833
  "trace-span-count": "trace-span-count";
5578
5834
  "trace-span-duration": "trace-span-duration";
5579
5835
  "search-rubric": "search-rubric";
5580
5836
  webhook: "webhook";
5581
5837
  "word-count": "word-count";
5582
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
5838
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
5583
5839
  "select-best": "select-best";
5584
5840
  human: "human";
5585
5841
  "max-score": "max-score";
@@ -5659,6 +5915,12 @@ declare const AtomicTestCaseSchema: z.ZodObject<{
5659
5915
  "similar:euclidean": "similar:euclidean";
5660
5916
  "starts-with": "starts-with";
5661
5917
  "tool-call-f1": "tool-call-f1";
5918
+ "skill-used": "skill-used";
5919
+ "trajectory:goal-success": "trajectory:goal-success";
5920
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
5921
+ "trajectory:step-count": "trajectory:step-count";
5922
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
5923
+ "trajectory:tool-used": "trajectory:tool-used";
5662
5924
  "trace-error-spans": "trace-error-spans";
5663
5925
  "trace-span-count": "trace-span-count";
5664
5926
  "trace-span-duration": "trace-span-duration";
@@ -5719,13 +5981,19 @@ declare const AtomicTestCaseSchema: z.ZodObject<{
5719
5981
  "similar:euclidean": "similar:euclidean";
5720
5982
  "starts-with": "starts-with";
5721
5983
  "tool-call-f1": "tool-call-f1";
5984
+ "skill-used": "skill-used";
5985
+ "trajectory:goal-success": "trajectory:goal-success";
5986
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
5987
+ "trajectory:step-count": "trajectory:step-count";
5988
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
5989
+ "trajectory:tool-used": "trajectory:tool-used";
5722
5990
  "trace-error-spans": "trace-error-spans";
5723
5991
  "trace-span-count": "trace-span-count";
5724
5992
  "trace-span-duration": "trace-span-duration";
5725
5993
  "search-rubric": "search-rubric";
5726
5994
  webhook: "webhook";
5727
5995
  "word-count": "word-count";
5728
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
5996
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
5729
5997
  "select-best": "select-best";
5730
5998
  human: "human";
5731
5999
  "max-score": "max-score";
@@ -5982,7 +6250,12 @@ declare const TestSuiteSchema: z.ZodObject<{
5982
6250
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
5983
6251
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
5984
6252
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
6253
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
6254
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
6255
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
5985
6256
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
6257
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
6258
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
5986
6259
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
5987
6260
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
5988
6261
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -6101,6 +6374,12 @@ declare const TestSuiteSchema: z.ZodObject<{
6101
6374
  "similar:euclidean": "similar:euclidean";
6102
6375
  "starts-with": "starts-with";
6103
6376
  "tool-call-f1": "tool-call-f1";
6377
+ "skill-used": "skill-used";
6378
+ "trajectory:goal-success": "trajectory:goal-success";
6379
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
6380
+ "trajectory:step-count": "trajectory:step-count";
6381
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
6382
+ "trajectory:tool-used": "trajectory:tool-used";
6104
6383
  "trace-error-spans": "trace-error-spans";
6105
6384
  "trace-span-count": "trace-span-count";
6106
6385
  "trace-span-duration": "trace-span-duration";
@@ -6161,13 +6440,19 @@ declare const TestSuiteSchema: z.ZodObject<{
6161
6440
  "similar:euclidean": "similar:euclidean";
6162
6441
  "starts-with": "starts-with";
6163
6442
  "tool-call-f1": "tool-call-f1";
6443
+ "skill-used": "skill-used";
6444
+ "trajectory:goal-success": "trajectory:goal-success";
6445
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
6446
+ "trajectory:step-count": "trajectory:step-count";
6447
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
6448
+ "trajectory:tool-used": "trajectory:tool-used";
6164
6449
  "trace-error-spans": "trace-error-spans";
6165
6450
  "trace-span-count": "trace-span-count";
6166
6451
  "trace-span-duration": "trace-span-duration";
6167
6452
  "search-rubric": "search-rubric";
6168
6453
  webhook: "webhook";
6169
6454
  "word-count": "word-count";
6170
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
6455
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
6171
6456
  "select-best": "select-best";
6172
6457
  human: "human";
6173
6458
  "max-score": "max-score";
@@ -6247,6 +6532,12 @@ declare const TestSuiteSchema: z.ZodObject<{
6247
6532
  "similar:euclidean": "similar:euclidean";
6248
6533
  "starts-with": "starts-with";
6249
6534
  "tool-call-f1": "tool-call-f1";
6535
+ "skill-used": "skill-used";
6536
+ "trajectory:goal-success": "trajectory:goal-success";
6537
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
6538
+ "trajectory:step-count": "trajectory:step-count";
6539
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
6540
+ "trajectory:tool-used": "trajectory:tool-used";
6250
6541
  "trace-error-spans": "trace-error-spans";
6251
6542
  "trace-span-count": "trace-span-count";
6252
6543
  "trace-span-duration": "trace-span-duration";
@@ -6307,13 +6598,19 @@ declare const TestSuiteSchema: z.ZodObject<{
6307
6598
  "similar:euclidean": "similar:euclidean";
6308
6599
  "starts-with": "starts-with";
6309
6600
  "tool-call-f1": "tool-call-f1";
6601
+ "skill-used": "skill-used";
6602
+ "trajectory:goal-success": "trajectory:goal-success";
6603
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
6604
+ "trajectory:step-count": "trajectory:step-count";
6605
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
6606
+ "trajectory:tool-used": "trajectory:tool-used";
6310
6607
  "trace-error-spans": "trace-error-spans";
6311
6608
  "trace-span-count": "trace-span-count";
6312
6609
  "trace-span-duration": "trace-span-duration";
6313
6610
  "search-rubric": "search-rubric";
6314
6611
  webhook: "webhook";
6315
6612
  "word-count": "word-count";
6316
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
6613
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
6317
6614
  "select-best": "select-best";
6318
6615
  human: "human";
6319
6616
  "max-score": "max-score";
@@ -6524,7 +6821,12 @@ declare const TestSuiteSchema: z.ZodObject<{
6524
6821
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
6525
6822
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
6526
6823
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
6824
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
6825
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
6826
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
6527
6827
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
6828
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
6829
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
6528
6830
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
6529
6831
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
6530
6832
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -6643,6 +6945,12 @@ declare const TestSuiteSchema: z.ZodObject<{
6643
6945
  "similar:euclidean": "similar:euclidean";
6644
6946
  "starts-with": "starts-with";
6645
6947
  "tool-call-f1": "tool-call-f1";
6948
+ "skill-used": "skill-used";
6949
+ "trajectory:goal-success": "trajectory:goal-success";
6950
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
6951
+ "trajectory:step-count": "trajectory:step-count";
6952
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
6953
+ "trajectory:tool-used": "trajectory:tool-used";
6646
6954
  "trace-error-spans": "trace-error-spans";
6647
6955
  "trace-span-count": "trace-span-count";
6648
6956
  "trace-span-duration": "trace-span-duration";
@@ -6703,13 +7011,19 @@ declare const TestSuiteSchema: z.ZodObject<{
6703
7011
  "similar:euclidean": "similar:euclidean";
6704
7012
  "starts-with": "starts-with";
6705
7013
  "tool-call-f1": "tool-call-f1";
7014
+ "skill-used": "skill-used";
7015
+ "trajectory:goal-success": "trajectory:goal-success";
7016
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
7017
+ "trajectory:step-count": "trajectory:step-count";
7018
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
7019
+ "trajectory:tool-used": "trajectory:tool-used";
6706
7020
  "trace-error-spans": "trace-error-spans";
6707
7021
  "trace-span-count": "trace-span-count";
6708
7022
  "trace-span-duration": "trace-span-duration";
6709
7023
  "search-rubric": "search-rubric";
6710
7024
  webhook: "webhook";
6711
7025
  "word-count": "word-count";
6712
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
7026
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
6713
7027
  "select-best": "select-best";
6714
7028
  human: "human";
6715
7029
  "max-score": "max-score";
@@ -6789,6 +7103,12 @@ declare const TestSuiteSchema: z.ZodObject<{
6789
7103
  "similar:euclidean": "similar:euclidean";
6790
7104
  "starts-with": "starts-with";
6791
7105
  "tool-call-f1": "tool-call-f1";
7106
+ "skill-used": "skill-used";
7107
+ "trajectory:goal-success": "trajectory:goal-success";
7108
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
7109
+ "trajectory:step-count": "trajectory:step-count";
7110
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
7111
+ "trajectory:tool-used": "trajectory:tool-used";
6792
7112
  "trace-error-spans": "trace-error-spans";
6793
7113
  "trace-span-count": "trace-span-count";
6794
7114
  "trace-span-duration": "trace-span-duration";
@@ -6849,13 +7169,19 @@ declare const TestSuiteSchema: z.ZodObject<{
6849
7169
  "similar:euclidean": "similar:euclidean";
6850
7170
  "starts-with": "starts-with";
6851
7171
  "tool-call-f1": "tool-call-f1";
7172
+ "skill-used": "skill-used";
7173
+ "trajectory:goal-success": "trajectory:goal-success";
7174
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
7175
+ "trajectory:step-count": "trajectory:step-count";
7176
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
7177
+ "trajectory:tool-used": "trajectory:tool-used";
6852
7178
  "trace-error-spans": "trace-error-spans";
6853
7179
  "trace-span-count": "trace-span-count";
6854
7180
  "trace-span-duration": "trace-span-duration";
6855
7181
  "search-rubric": "search-rubric";
6856
7182
  webhook: "webhook";
6857
7183
  "word-count": "word-count";
6858
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
7184
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
6859
7185
  "select-best": "select-best";
6860
7186
  human: "human";
6861
7187
  "max-score": "max-score";
@@ -7064,7 +7390,12 @@ declare const TestSuiteSchema: z.ZodObject<{
7064
7390
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
7065
7391
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
7066
7392
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
7393
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
7394
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
7395
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
7067
7396
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
7397
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
7398
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
7068
7399
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
7069
7400
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
7070
7401
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -7183,6 +7514,12 @@ declare const TestSuiteSchema: z.ZodObject<{
7183
7514
  "similar:euclidean": "similar:euclidean";
7184
7515
  "starts-with": "starts-with";
7185
7516
  "tool-call-f1": "tool-call-f1";
7517
+ "skill-used": "skill-used";
7518
+ "trajectory:goal-success": "trajectory:goal-success";
7519
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
7520
+ "trajectory:step-count": "trajectory:step-count";
7521
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
7522
+ "trajectory:tool-used": "trajectory:tool-used";
7186
7523
  "trace-error-spans": "trace-error-spans";
7187
7524
  "trace-span-count": "trace-span-count";
7188
7525
  "trace-span-duration": "trace-span-duration";
@@ -7243,13 +7580,19 @@ declare const TestSuiteSchema: z.ZodObject<{
7243
7580
  "similar:euclidean": "similar:euclidean";
7244
7581
  "starts-with": "starts-with";
7245
7582
  "tool-call-f1": "tool-call-f1";
7583
+ "skill-used": "skill-used";
7584
+ "trajectory:goal-success": "trajectory:goal-success";
7585
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
7586
+ "trajectory:step-count": "trajectory:step-count";
7587
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
7588
+ "trajectory:tool-used": "trajectory:tool-used";
7246
7589
  "trace-error-spans": "trace-error-spans";
7247
7590
  "trace-span-count": "trace-span-count";
7248
7591
  "trace-span-duration": "trace-span-duration";
7249
7592
  "search-rubric": "search-rubric";
7250
7593
  webhook: "webhook";
7251
7594
  "word-count": "word-count";
7252
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
7595
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
7253
7596
  "select-best": "select-best";
7254
7597
  human: "human";
7255
7598
  "max-score": "max-score";
@@ -7329,6 +7672,12 @@ declare const TestSuiteSchema: z.ZodObject<{
7329
7672
  "similar:euclidean": "similar:euclidean";
7330
7673
  "starts-with": "starts-with";
7331
7674
  "tool-call-f1": "tool-call-f1";
7675
+ "skill-used": "skill-used";
7676
+ "trajectory:goal-success": "trajectory:goal-success";
7677
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
7678
+ "trajectory:step-count": "trajectory:step-count";
7679
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
7680
+ "trajectory:tool-used": "trajectory:tool-used";
7332
7681
  "trace-error-spans": "trace-error-spans";
7333
7682
  "trace-span-count": "trace-span-count";
7334
7683
  "trace-span-duration": "trace-span-duration";
@@ -7389,13 +7738,19 @@ declare const TestSuiteSchema: z.ZodObject<{
7389
7738
  "similar:euclidean": "similar:euclidean";
7390
7739
  "starts-with": "starts-with";
7391
7740
  "tool-call-f1": "tool-call-f1";
7741
+ "skill-used": "skill-used";
7742
+ "trajectory:goal-success": "trajectory:goal-success";
7743
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
7744
+ "trajectory:step-count": "trajectory:step-count";
7745
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
7746
+ "trajectory:tool-used": "trajectory:tool-used";
7392
7747
  "trace-error-spans": "trace-error-spans";
7393
7748
  "trace-span-count": "trace-span-count";
7394
7749
  "trace-span-duration": "trace-span-duration";
7395
7750
  "search-rubric": "search-rubric";
7396
7751
  webhook: "webhook";
7397
7752
  "word-count": "word-count";
7398
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
7753
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
7399
7754
  "select-best": "select-best";
7400
7755
  human: "human";
7401
7756
  "max-score": "max-score";
@@ -7606,7 +7961,12 @@ declare const TestSuiteSchema: z.ZodObject<{
7606
7961
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
7607
7962
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
7608
7963
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
7964
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
7965
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
7966
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
7609
7967
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
7968
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
7969
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
7610
7970
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
7611
7971
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
7612
7972
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -7723,6 +8083,12 @@ declare const TestSuiteSchema: z.ZodObject<{
7723
8083
  "similar:euclidean": "similar:euclidean";
7724
8084
  "starts-with": "starts-with";
7725
8085
  "tool-call-f1": "tool-call-f1";
8086
+ "skill-used": "skill-used";
8087
+ "trajectory:goal-success": "trajectory:goal-success";
8088
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
8089
+ "trajectory:step-count": "trajectory:step-count";
8090
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
8091
+ "trajectory:tool-used": "trajectory:tool-used";
7726
8092
  "trace-error-spans": "trace-error-spans";
7727
8093
  "trace-span-count": "trace-span-count";
7728
8094
  "trace-span-duration": "trace-span-duration";
@@ -7783,13 +8149,19 @@ declare const TestSuiteSchema: z.ZodObject<{
7783
8149
  "similar:euclidean": "similar:euclidean";
7784
8150
  "starts-with": "starts-with";
7785
8151
  "tool-call-f1": "tool-call-f1";
8152
+ "skill-used": "skill-used";
8153
+ "trajectory:goal-success": "trajectory:goal-success";
8154
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
8155
+ "trajectory:step-count": "trajectory:step-count";
8156
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
8157
+ "trajectory:tool-used": "trajectory:tool-used";
7786
8158
  "trace-error-spans": "trace-error-spans";
7787
8159
  "trace-span-count": "trace-span-count";
7788
8160
  "trace-span-duration": "trace-span-duration";
7789
8161
  "search-rubric": "search-rubric";
7790
8162
  webhook: "webhook";
7791
8163
  "word-count": "word-count";
7792
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
8164
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
7793
8165
  "select-best": "select-best";
7794
8166
  human: "human";
7795
8167
  "max-score": "max-score";
@@ -7869,6 +8241,12 @@ declare const TestSuiteSchema: z.ZodObject<{
7869
8241
  "similar:euclidean": "similar:euclidean";
7870
8242
  "starts-with": "starts-with";
7871
8243
  "tool-call-f1": "tool-call-f1";
8244
+ "skill-used": "skill-used";
8245
+ "trajectory:goal-success": "trajectory:goal-success";
8246
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
8247
+ "trajectory:step-count": "trajectory:step-count";
8248
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
8249
+ "trajectory:tool-used": "trajectory:tool-used";
7872
8250
  "trace-error-spans": "trace-error-spans";
7873
8251
  "trace-span-count": "trace-span-count";
7874
8252
  "trace-span-duration": "trace-span-duration";
@@ -7929,13 +8307,19 @@ declare const TestSuiteSchema: z.ZodObject<{
7929
8307
  "similar:euclidean": "similar:euclidean";
7930
8308
  "starts-with": "starts-with";
7931
8309
  "tool-call-f1": "tool-call-f1";
8310
+ "skill-used": "skill-used";
8311
+ "trajectory:goal-success": "trajectory:goal-success";
8312
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
8313
+ "trajectory:step-count": "trajectory:step-count";
8314
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
8315
+ "trajectory:tool-used": "trajectory:tool-used";
7932
8316
  "trace-error-spans": "trace-error-spans";
7933
8317
  "trace-span-count": "trace-span-count";
7934
8318
  "trace-span-duration": "trace-span-duration";
7935
8319
  "search-rubric": "search-rubric";
7936
8320
  webhook: "webhook";
7937
8321
  "word-count": "word-count";
7938
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
8322
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
7939
8323
  "select-best": "select-best";
7940
8324
  human: "human";
7941
8325
  "max-score": "max-score";
@@ -8135,7 +8519,12 @@ declare const TestSuiteSchema: z.ZodObject<{
8135
8519
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
8136
8520
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
8137
8521
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
8522
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
8523
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
8524
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
8138
8525
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
8526
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
8527
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
8139
8528
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
8140
8529
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
8141
8530
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -8304,7 +8693,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
8304
8693
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
8305
8694
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
8306
8695
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
8696
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
8697
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
8698
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
8307
8699
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
8700
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
8701
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
8308
8702
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
8309
8703
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
8310
8704
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -8431,7 +8825,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
8431
8825
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
8432
8826
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
8433
8827
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
8828
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
8829
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
8830
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
8434
8831
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
8832
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
8833
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
8435
8834
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
8436
8835
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
8437
8836
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -8574,7 +8973,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
8574
8973
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
8575
8974
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
8576
8975
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
8976
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
8977
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
8978
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
8577
8979
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
8980
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
8981
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
8578
8982
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
8579
8983
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
8580
8984
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -8693,6 +9097,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
8693
9097
  "similar:euclidean": "similar:euclidean";
8694
9098
  "starts-with": "starts-with";
8695
9099
  "tool-call-f1": "tool-call-f1";
9100
+ "skill-used": "skill-used";
9101
+ "trajectory:goal-success": "trajectory:goal-success";
9102
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
9103
+ "trajectory:step-count": "trajectory:step-count";
9104
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
9105
+ "trajectory:tool-used": "trajectory:tool-used";
8696
9106
  "trace-error-spans": "trace-error-spans";
8697
9107
  "trace-span-count": "trace-span-count";
8698
9108
  "trace-span-duration": "trace-span-duration";
@@ -8753,13 +9163,19 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
8753
9163
  "similar:euclidean": "similar:euclidean";
8754
9164
  "starts-with": "starts-with";
8755
9165
  "tool-call-f1": "tool-call-f1";
9166
+ "skill-used": "skill-used";
9167
+ "trajectory:goal-success": "trajectory:goal-success";
9168
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
9169
+ "trajectory:step-count": "trajectory:step-count";
9170
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
9171
+ "trajectory:tool-used": "trajectory:tool-used";
8756
9172
  "trace-error-spans": "trace-error-spans";
8757
9173
  "trace-span-count": "trace-span-count";
8758
9174
  "trace-span-duration": "trace-span-duration";
8759
9175
  "search-rubric": "search-rubric";
8760
9176
  webhook: "webhook";
8761
9177
  "word-count": "word-count";
8762
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
9178
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
8763
9179
  "select-best": "select-best";
8764
9180
  human: "human";
8765
9181
  "max-score": "max-score";
@@ -8839,6 +9255,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
8839
9255
  "similar:euclidean": "similar:euclidean";
8840
9256
  "starts-with": "starts-with";
8841
9257
  "tool-call-f1": "tool-call-f1";
9258
+ "skill-used": "skill-used";
9259
+ "trajectory:goal-success": "trajectory:goal-success";
9260
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
9261
+ "trajectory:step-count": "trajectory:step-count";
9262
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
9263
+ "trajectory:tool-used": "trajectory:tool-used";
8842
9264
  "trace-error-spans": "trace-error-spans";
8843
9265
  "trace-span-count": "trace-span-count";
8844
9266
  "trace-span-duration": "trace-span-duration";
@@ -8899,13 +9321,19 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
8899
9321
  "similar:euclidean": "similar:euclidean";
8900
9322
  "starts-with": "starts-with";
8901
9323
  "tool-call-f1": "tool-call-f1";
9324
+ "skill-used": "skill-used";
9325
+ "trajectory:goal-success": "trajectory:goal-success";
9326
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
9327
+ "trajectory:step-count": "trajectory:step-count";
9328
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
9329
+ "trajectory:tool-used": "trajectory:tool-used";
8902
9330
  "trace-error-spans": "trace-error-spans";
8903
9331
  "trace-span-count": "trace-span-count";
8904
9332
  "trace-span-duration": "trace-span-duration";
8905
9333
  "search-rubric": "search-rubric";
8906
9334
  webhook: "webhook";
8907
9335
  "word-count": "word-count";
8908
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
9336
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
8909
9337
  "select-best": "select-best";
8910
9338
  human: "human";
8911
9339
  "max-score": "max-score";
@@ -9122,7 +9550,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
9122
9550
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
9123
9551
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
9124
9552
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
9553
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
9554
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
9555
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
9125
9556
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
9557
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
9558
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
9126
9559
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
9127
9560
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
9128
9561
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -9241,6 +9674,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
9241
9674
  "similar:euclidean": "similar:euclidean";
9242
9675
  "starts-with": "starts-with";
9243
9676
  "tool-call-f1": "tool-call-f1";
9677
+ "skill-used": "skill-used";
9678
+ "trajectory:goal-success": "trajectory:goal-success";
9679
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
9680
+ "trajectory:step-count": "trajectory:step-count";
9681
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
9682
+ "trajectory:tool-used": "trajectory:tool-used";
9244
9683
  "trace-error-spans": "trace-error-spans";
9245
9684
  "trace-span-count": "trace-span-count";
9246
9685
  "trace-span-duration": "trace-span-duration";
@@ -9301,13 +9740,19 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
9301
9740
  "similar:euclidean": "similar:euclidean";
9302
9741
  "starts-with": "starts-with";
9303
9742
  "tool-call-f1": "tool-call-f1";
9743
+ "skill-used": "skill-used";
9744
+ "trajectory:goal-success": "trajectory:goal-success";
9745
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
9746
+ "trajectory:step-count": "trajectory:step-count";
9747
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
9748
+ "trajectory:tool-used": "trajectory:tool-used";
9304
9749
  "trace-error-spans": "trace-error-spans";
9305
9750
  "trace-span-count": "trace-span-count";
9306
9751
  "trace-span-duration": "trace-span-duration";
9307
9752
  "search-rubric": "search-rubric";
9308
9753
  webhook: "webhook";
9309
9754
  "word-count": "word-count";
9310
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
9755
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
9311
9756
  "select-best": "select-best";
9312
9757
  human: "human";
9313
9758
  "max-score": "max-score";
@@ -9387,6 +9832,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
9387
9832
  "similar:euclidean": "similar:euclidean";
9388
9833
  "starts-with": "starts-with";
9389
9834
  "tool-call-f1": "tool-call-f1";
9835
+ "skill-used": "skill-used";
9836
+ "trajectory:goal-success": "trajectory:goal-success";
9837
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
9838
+ "trajectory:step-count": "trajectory:step-count";
9839
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
9840
+ "trajectory:tool-used": "trajectory:tool-used";
9390
9841
  "trace-error-spans": "trace-error-spans";
9391
9842
  "trace-span-count": "trace-span-count";
9392
9843
  "trace-span-duration": "trace-span-duration";
@@ -9447,13 +9898,19 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
9447
9898
  "similar:euclidean": "similar:euclidean";
9448
9899
  "starts-with": "starts-with";
9449
9900
  "tool-call-f1": "tool-call-f1";
9901
+ "skill-used": "skill-used";
9902
+ "trajectory:goal-success": "trajectory:goal-success";
9903
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
9904
+ "trajectory:step-count": "trajectory:step-count";
9905
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
9906
+ "trajectory:tool-used": "trajectory:tool-used";
9450
9907
  "trace-error-spans": "trace-error-spans";
9451
9908
  "trace-span-count": "trace-span-count";
9452
9909
  "trace-span-duration": "trace-span-duration";
9453
9910
  "search-rubric": "search-rubric";
9454
9911
  webhook: "webhook";
9455
9912
  "word-count": "word-count";
9456
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
9913
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
9457
9914
  "select-best": "select-best";
9458
9915
  human: "human";
9459
9916
  "max-score": "max-score";
@@ -9662,7 +10119,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
9662
10119
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
9663
10120
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
9664
10121
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
10122
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
10123
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
10124
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
9665
10125
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
10126
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
10127
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
9666
10128
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
9667
10129
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
9668
10130
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -9781,6 +10243,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
9781
10243
  "similar:euclidean": "similar:euclidean";
9782
10244
  "starts-with": "starts-with";
9783
10245
  "tool-call-f1": "tool-call-f1";
10246
+ "skill-used": "skill-used";
10247
+ "trajectory:goal-success": "trajectory:goal-success";
10248
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
10249
+ "trajectory:step-count": "trajectory:step-count";
10250
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
10251
+ "trajectory:tool-used": "trajectory:tool-used";
9784
10252
  "trace-error-spans": "trace-error-spans";
9785
10253
  "trace-span-count": "trace-span-count";
9786
10254
  "trace-span-duration": "trace-span-duration";
@@ -9841,13 +10309,19 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
9841
10309
  "similar:euclidean": "similar:euclidean";
9842
10310
  "starts-with": "starts-with";
9843
10311
  "tool-call-f1": "tool-call-f1";
10312
+ "skill-used": "skill-used";
10313
+ "trajectory:goal-success": "trajectory:goal-success";
10314
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
10315
+ "trajectory:step-count": "trajectory:step-count";
10316
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
10317
+ "trajectory:tool-used": "trajectory:tool-used";
9844
10318
  "trace-error-spans": "trace-error-spans";
9845
10319
  "trace-span-count": "trace-span-count";
9846
10320
  "trace-span-duration": "trace-span-duration";
9847
10321
  "search-rubric": "search-rubric";
9848
10322
  webhook: "webhook";
9849
10323
  "word-count": "word-count";
9850
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
10324
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
9851
10325
  "select-best": "select-best";
9852
10326
  human: "human";
9853
10327
  "max-score": "max-score";
@@ -9927,6 +10401,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
9927
10401
  "similar:euclidean": "similar:euclidean";
9928
10402
  "starts-with": "starts-with";
9929
10403
  "tool-call-f1": "tool-call-f1";
10404
+ "skill-used": "skill-used";
10405
+ "trajectory:goal-success": "trajectory:goal-success";
10406
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
10407
+ "trajectory:step-count": "trajectory:step-count";
10408
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
10409
+ "trajectory:tool-used": "trajectory:tool-used";
9930
10410
  "trace-error-spans": "trace-error-spans";
9931
10411
  "trace-span-count": "trace-span-count";
9932
10412
  "trace-span-duration": "trace-span-duration";
@@ -9987,13 +10467,19 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
9987
10467
  "similar:euclidean": "similar:euclidean";
9988
10468
  "starts-with": "starts-with";
9989
10469
  "tool-call-f1": "tool-call-f1";
10470
+ "skill-used": "skill-used";
10471
+ "trajectory:goal-success": "trajectory:goal-success";
10472
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
10473
+ "trajectory:step-count": "trajectory:step-count";
10474
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
10475
+ "trajectory:tool-used": "trajectory:tool-used";
9990
10476
  "trace-error-spans": "trace-error-spans";
9991
10477
  "trace-span-count": "trace-span-count";
9992
10478
  "trace-span-duration": "trace-span-duration";
9993
10479
  "search-rubric": "search-rubric";
9994
10480
  webhook: "webhook";
9995
10481
  "word-count": "word-count";
9996
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
10482
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
9997
10483
  "select-best": "select-best";
9998
10484
  human: "human";
9999
10485
  "max-score": "max-score";
@@ -10204,7 +10690,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
10204
10690
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
10205
10691
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
10206
10692
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
10693
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
10694
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
10695
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
10207
10696
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
10697
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
10698
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
10208
10699
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
10209
10700
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
10210
10701
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -10321,6 +10812,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
10321
10812
  "similar:euclidean": "similar:euclidean";
10322
10813
  "starts-with": "starts-with";
10323
10814
  "tool-call-f1": "tool-call-f1";
10815
+ "skill-used": "skill-used";
10816
+ "trajectory:goal-success": "trajectory:goal-success";
10817
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
10818
+ "trajectory:step-count": "trajectory:step-count";
10819
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
10820
+ "trajectory:tool-used": "trajectory:tool-used";
10324
10821
  "trace-error-spans": "trace-error-spans";
10325
10822
  "trace-span-count": "trace-span-count";
10326
10823
  "trace-span-duration": "trace-span-duration";
@@ -10381,13 +10878,19 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
10381
10878
  "similar:euclidean": "similar:euclidean";
10382
10879
  "starts-with": "starts-with";
10383
10880
  "tool-call-f1": "tool-call-f1";
10881
+ "skill-used": "skill-used";
10882
+ "trajectory:goal-success": "trajectory:goal-success";
10883
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
10884
+ "trajectory:step-count": "trajectory:step-count";
10885
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
10886
+ "trajectory:tool-used": "trajectory:tool-used";
10384
10887
  "trace-error-spans": "trace-error-spans";
10385
10888
  "trace-span-count": "trace-span-count";
10386
10889
  "trace-span-duration": "trace-span-duration";
10387
10890
  "search-rubric": "search-rubric";
10388
10891
  webhook: "webhook";
10389
10892
  "word-count": "word-count";
10390
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
10893
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
10391
10894
  "select-best": "select-best";
10392
10895
  human: "human";
10393
10896
  "max-score": "max-score";
@@ -10467,6 +10970,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
10467
10970
  "similar:euclidean": "similar:euclidean";
10468
10971
  "starts-with": "starts-with";
10469
10972
  "tool-call-f1": "tool-call-f1";
10973
+ "skill-used": "skill-used";
10974
+ "trajectory:goal-success": "trajectory:goal-success";
10975
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
10976
+ "trajectory:step-count": "trajectory:step-count";
10977
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
10978
+ "trajectory:tool-used": "trajectory:tool-used";
10470
10979
  "trace-error-spans": "trace-error-spans";
10471
10980
  "trace-span-count": "trace-span-count";
10472
10981
  "trace-span-duration": "trace-span-duration";
@@ -10527,13 +11036,19 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
10527
11036
  "similar:euclidean": "similar:euclidean";
10528
11037
  "starts-with": "starts-with";
10529
11038
  "tool-call-f1": "tool-call-f1";
11039
+ "skill-used": "skill-used";
11040
+ "trajectory:goal-success": "trajectory:goal-success";
11041
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
11042
+ "trajectory:step-count": "trajectory:step-count";
11043
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
11044
+ "trajectory:tool-used": "trajectory:tool-used";
10530
11045
  "trace-error-spans": "trace-error-spans";
10531
11046
  "trace-span-count": "trace-span-count";
10532
11047
  "trace-span-duration": "trace-span-duration";
10533
11048
  "search-rubric": "search-rubric";
10534
11049
  webhook: "webhook";
10535
11050
  "word-count": "word-count";
10536
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
11051
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
10537
11052
  "select-best": "select-best";
10538
11053
  human: "human";
10539
11054
  "max-score": "max-score";
@@ -10738,7 +11253,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
10738
11253
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
10739
11254
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
10740
11255
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
11256
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
11257
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
11258
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
10741
11259
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
11260
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
11261
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
10742
11262
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
10743
11263
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
10744
11264
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -10884,7 +11404,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
10884
11404
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
10885
11405
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
10886
11406
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
11407
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
11408
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
11409
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
10887
11410
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
11411
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
11412
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
10888
11413
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
10889
11414
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
10890
11415
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -11079,7 +11604,12 @@ declare const TestSuiteConfigSchema: z.ZodObject<{
11079
11604
  OPENAI_API_KEY?: string | undefined;
11080
11605
  OPENAI_BASE_URL?: string | undefined;
11081
11606
  OPENAI_ORGANIZATION?: string | undefined;
11607
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
11608
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
11609
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
11082
11610
  CODEX_API_KEY?: string | undefined;
11611
+ OPENCLAW_CONFIG_PATH?: string | undefined;
11612
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
11083
11613
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
11084
11614
  OPENCLAW_GATEWAY_URL?: string | undefined;
11085
11615
  PALM_API_HOST?: string | undefined;
@@ -11287,7 +11817,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
11287
11817
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
11288
11818
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
11289
11819
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
11820
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
11821
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
11822
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
11290
11823
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
11824
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
11825
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
11291
11826
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
11292
11827
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
11293
11828
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -11406,6 +11941,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
11406
11941
  "similar:euclidean": "similar:euclidean";
11407
11942
  "starts-with": "starts-with";
11408
11943
  "tool-call-f1": "tool-call-f1";
11944
+ "skill-used": "skill-used";
11945
+ "trajectory:goal-success": "trajectory:goal-success";
11946
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
11947
+ "trajectory:step-count": "trajectory:step-count";
11948
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
11949
+ "trajectory:tool-used": "trajectory:tool-used";
11409
11950
  "trace-error-spans": "trace-error-spans";
11410
11951
  "trace-span-count": "trace-span-count";
11411
11952
  "trace-span-duration": "trace-span-duration";
@@ -11466,13 +12007,19 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
11466
12007
  "similar:euclidean": "similar:euclidean";
11467
12008
  "starts-with": "starts-with";
11468
12009
  "tool-call-f1": "tool-call-f1";
12010
+ "skill-used": "skill-used";
12011
+ "trajectory:goal-success": "trajectory:goal-success";
12012
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
12013
+ "trajectory:step-count": "trajectory:step-count";
12014
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
12015
+ "trajectory:tool-used": "trajectory:tool-used";
11469
12016
  "trace-error-spans": "trace-error-spans";
11470
12017
  "trace-span-count": "trace-span-count";
11471
12018
  "trace-span-duration": "trace-span-duration";
11472
12019
  "search-rubric": "search-rubric";
11473
12020
  webhook: "webhook";
11474
12021
  "word-count": "word-count";
11475
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
12022
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
11476
12023
  "select-best": "select-best";
11477
12024
  human: "human";
11478
12025
  "max-score": "max-score";
@@ -11552,6 +12099,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
11552
12099
  "similar:euclidean": "similar:euclidean";
11553
12100
  "starts-with": "starts-with";
11554
12101
  "tool-call-f1": "tool-call-f1";
12102
+ "skill-used": "skill-used";
12103
+ "trajectory:goal-success": "trajectory:goal-success";
12104
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
12105
+ "trajectory:step-count": "trajectory:step-count";
12106
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
12107
+ "trajectory:tool-used": "trajectory:tool-used";
11555
12108
  "trace-error-spans": "trace-error-spans";
11556
12109
  "trace-span-count": "trace-span-count";
11557
12110
  "trace-span-duration": "trace-span-duration";
@@ -11612,13 +12165,19 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
11612
12165
  "similar:euclidean": "similar:euclidean";
11613
12166
  "starts-with": "starts-with";
11614
12167
  "tool-call-f1": "tool-call-f1";
12168
+ "skill-used": "skill-used";
12169
+ "trajectory:goal-success": "trajectory:goal-success";
12170
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
12171
+ "trajectory:step-count": "trajectory:step-count";
12172
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
12173
+ "trajectory:tool-used": "trajectory:tool-used";
11615
12174
  "trace-error-spans": "trace-error-spans";
11616
12175
  "trace-span-count": "trace-span-count";
11617
12176
  "trace-span-duration": "trace-span-duration";
11618
12177
  "search-rubric": "search-rubric";
11619
12178
  webhook: "webhook";
11620
12179
  "word-count": "word-count";
11621
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
12180
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
11622
12181
  "select-best": "select-best";
11623
12182
  human: "human";
11624
12183
  "max-score": "max-score";
@@ -11835,7 +12394,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
11835
12394
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
11836
12395
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
11837
12396
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
12397
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
12398
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
12399
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
11838
12400
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
12401
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
12402
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
11839
12403
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
11840
12404
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
11841
12405
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -11954,6 +12518,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
11954
12518
  "similar:euclidean": "similar:euclidean";
11955
12519
  "starts-with": "starts-with";
11956
12520
  "tool-call-f1": "tool-call-f1";
12521
+ "skill-used": "skill-used";
12522
+ "trajectory:goal-success": "trajectory:goal-success";
12523
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
12524
+ "trajectory:step-count": "trajectory:step-count";
12525
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
12526
+ "trajectory:tool-used": "trajectory:tool-used";
11957
12527
  "trace-error-spans": "trace-error-spans";
11958
12528
  "trace-span-count": "trace-span-count";
11959
12529
  "trace-span-duration": "trace-span-duration";
@@ -12014,13 +12584,19 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
12014
12584
  "similar:euclidean": "similar:euclidean";
12015
12585
  "starts-with": "starts-with";
12016
12586
  "tool-call-f1": "tool-call-f1";
12587
+ "skill-used": "skill-used";
12588
+ "trajectory:goal-success": "trajectory:goal-success";
12589
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
12590
+ "trajectory:step-count": "trajectory:step-count";
12591
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
12592
+ "trajectory:tool-used": "trajectory:tool-used";
12017
12593
  "trace-error-spans": "trace-error-spans";
12018
12594
  "trace-span-count": "trace-span-count";
12019
12595
  "trace-span-duration": "trace-span-duration";
12020
12596
  "search-rubric": "search-rubric";
12021
12597
  webhook: "webhook";
12022
12598
  "word-count": "word-count";
12023
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
12599
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
12024
12600
  "select-best": "select-best";
12025
12601
  human: "human";
12026
12602
  "max-score": "max-score";
@@ -12100,6 +12676,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
12100
12676
  "similar:euclidean": "similar:euclidean";
12101
12677
  "starts-with": "starts-with";
12102
12678
  "tool-call-f1": "tool-call-f1";
12679
+ "skill-used": "skill-used";
12680
+ "trajectory:goal-success": "trajectory:goal-success";
12681
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
12682
+ "trajectory:step-count": "trajectory:step-count";
12683
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
12684
+ "trajectory:tool-used": "trajectory:tool-used";
12103
12685
  "trace-error-spans": "trace-error-spans";
12104
12686
  "trace-span-count": "trace-span-count";
12105
12687
  "trace-span-duration": "trace-span-duration";
@@ -12160,13 +12742,19 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
12160
12742
  "similar:euclidean": "similar:euclidean";
12161
12743
  "starts-with": "starts-with";
12162
12744
  "tool-call-f1": "tool-call-f1";
12745
+ "skill-used": "skill-used";
12746
+ "trajectory:goal-success": "trajectory:goal-success";
12747
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
12748
+ "trajectory:step-count": "trajectory:step-count";
12749
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
12750
+ "trajectory:tool-used": "trajectory:tool-used";
12163
12751
  "trace-error-spans": "trace-error-spans";
12164
12752
  "trace-span-count": "trace-span-count";
12165
12753
  "trace-span-duration": "trace-span-duration";
12166
12754
  "search-rubric": "search-rubric";
12167
12755
  webhook: "webhook";
12168
12756
  "word-count": "word-count";
12169
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
12757
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
12170
12758
  "select-best": "select-best";
12171
12759
  human: "human";
12172
12760
  "max-score": "max-score";
@@ -12375,7 +12963,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
12375
12963
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
12376
12964
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
12377
12965
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
12966
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
12967
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
12968
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
12378
12969
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
12970
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
12971
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
12379
12972
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
12380
12973
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
12381
12974
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -12494,6 +13087,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
12494
13087
  "similar:euclidean": "similar:euclidean";
12495
13088
  "starts-with": "starts-with";
12496
13089
  "tool-call-f1": "tool-call-f1";
13090
+ "skill-used": "skill-used";
13091
+ "trajectory:goal-success": "trajectory:goal-success";
13092
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
13093
+ "trajectory:step-count": "trajectory:step-count";
13094
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
13095
+ "trajectory:tool-used": "trajectory:tool-used";
12497
13096
  "trace-error-spans": "trace-error-spans";
12498
13097
  "trace-span-count": "trace-span-count";
12499
13098
  "trace-span-duration": "trace-span-duration";
@@ -12554,13 +13153,19 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
12554
13153
  "similar:euclidean": "similar:euclidean";
12555
13154
  "starts-with": "starts-with";
12556
13155
  "tool-call-f1": "tool-call-f1";
13156
+ "skill-used": "skill-used";
13157
+ "trajectory:goal-success": "trajectory:goal-success";
13158
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
13159
+ "trajectory:step-count": "trajectory:step-count";
13160
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
13161
+ "trajectory:tool-used": "trajectory:tool-used";
12557
13162
  "trace-error-spans": "trace-error-spans";
12558
13163
  "trace-span-count": "trace-span-count";
12559
13164
  "trace-span-duration": "trace-span-duration";
12560
13165
  "search-rubric": "search-rubric";
12561
13166
  webhook: "webhook";
12562
13167
  "word-count": "word-count";
12563
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
13168
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
12564
13169
  "select-best": "select-best";
12565
13170
  human: "human";
12566
13171
  "max-score": "max-score";
@@ -12640,6 +13245,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
12640
13245
  "similar:euclidean": "similar:euclidean";
12641
13246
  "starts-with": "starts-with";
12642
13247
  "tool-call-f1": "tool-call-f1";
13248
+ "skill-used": "skill-used";
13249
+ "trajectory:goal-success": "trajectory:goal-success";
13250
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
13251
+ "trajectory:step-count": "trajectory:step-count";
13252
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
13253
+ "trajectory:tool-used": "trajectory:tool-used";
12643
13254
  "trace-error-spans": "trace-error-spans";
12644
13255
  "trace-span-count": "trace-span-count";
12645
13256
  "trace-span-duration": "trace-span-duration";
@@ -12700,13 +13311,19 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
12700
13311
  "similar:euclidean": "similar:euclidean";
12701
13312
  "starts-with": "starts-with";
12702
13313
  "tool-call-f1": "tool-call-f1";
13314
+ "skill-used": "skill-used";
13315
+ "trajectory:goal-success": "trajectory:goal-success";
13316
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
13317
+ "trajectory:step-count": "trajectory:step-count";
13318
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
13319
+ "trajectory:tool-used": "trajectory:tool-used";
12703
13320
  "trace-error-spans": "trace-error-spans";
12704
13321
  "trace-span-count": "trace-span-count";
12705
13322
  "trace-span-duration": "trace-span-duration";
12706
13323
  "search-rubric": "search-rubric";
12707
13324
  webhook: "webhook";
12708
13325
  "word-count": "word-count";
12709
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
13326
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
12710
13327
  "select-best": "select-best";
12711
13328
  human: "human";
12712
13329
  "max-score": "max-score";
@@ -12917,7 +13534,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
12917
13534
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
12918
13535
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
12919
13536
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
13537
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
13538
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
13539
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
12920
13540
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
13541
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
13542
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
12921
13543
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
12922
13544
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
12923
13545
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -13034,6 +13656,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
13034
13656
  "similar:euclidean": "similar:euclidean";
13035
13657
  "starts-with": "starts-with";
13036
13658
  "tool-call-f1": "tool-call-f1";
13659
+ "skill-used": "skill-used";
13660
+ "trajectory:goal-success": "trajectory:goal-success";
13661
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
13662
+ "trajectory:step-count": "trajectory:step-count";
13663
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
13664
+ "trajectory:tool-used": "trajectory:tool-used";
13037
13665
  "trace-error-spans": "trace-error-spans";
13038
13666
  "trace-span-count": "trace-span-count";
13039
13667
  "trace-span-duration": "trace-span-duration";
@@ -13094,13 +13722,19 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
13094
13722
  "similar:euclidean": "similar:euclidean";
13095
13723
  "starts-with": "starts-with";
13096
13724
  "tool-call-f1": "tool-call-f1";
13725
+ "skill-used": "skill-used";
13726
+ "trajectory:goal-success": "trajectory:goal-success";
13727
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
13728
+ "trajectory:step-count": "trajectory:step-count";
13729
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
13730
+ "trajectory:tool-used": "trajectory:tool-used";
13097
13731
  "trace-error-spans": "trace-error-spans";
13098
13732
  "trace-span-count": "trace-span-count";
13099
13733
  "trace-span-duration": "trace-span-duration";
13100
13734
  "search-rubric": "search-rubric";
13101
13735
  webhook: "webhook";
13102
13736
  "word-count": "word-count";
13103
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
13737
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
13104
13738
  "select-best": "select-best";
13105
13739
  human: "human";
13106
13740
  "max-score": "max-score";
@@ -13180,6 +13814,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
13180
13814
  "similar:euclidean": "similar:euclidean";
13181
13815
  "starts-with": "starts-with";
13182
13816
  "tool-call-f1": "tool-call-f1";
13817
+ "skill-used": "skill-used";
13818
+ "trajectory:goal-success": "trajectory:goal-success";
13819
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
13820
+ "trajectory:step-count": "trajectory:step-count";
13821
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
13822
+ "trajectory:tool-used": "trajectory:tool-used";
13183
13823
  "trace-error-spans": "trace-error-spans";
13184
13824
  "trace-span-count": "trace-span-count";
13185
13825
  "trace-span-duration": "trace-span-duration";
@@ -13240,13 +13880,19 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
13240
13880
  "similar:euclidean": "similar:euclidean";
13241
13881
  "starts-with": "starts-with";
13242
13882
  "tool-call-f1": "tool-call-f1";
13883
+ "skill-used": "skill-used";
13884
+ "trajectory:goal-success": "trajectory:goal-success";
13885
+ "trajectory:tool-args-match": "trajectory:tool-args-match";
13886
+ "trajectory:step-count": "trajectory:step-count";
13887
+ "trajectory:tool-sequence": "trajectory:tool-sequence";
13888
+ "trajectory:tool-used": "trajectory:tool-used";
13243
13889
  "trace-error-spans": "trace-error-spans";
13244
13890
  "trace-span-count": "trace-span-count";
13245
13891
  "trace-span-duration": "trace-span-duration";
13246
13892
  "search-rubric": "search-rubric";
13247
13893
  webhook: "webhook";
13248
13894
  "word-count": "word-count";
13249
- }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
13895
+ }>, z.ZodTransform<"not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count", "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count">>, z.ZodEnum<{
13250
13896
  "select-best": "select-best";
13251
13897
  human: "human";
13252
13898
  "max-score": "max-score";
@@ -13451,7 +14097,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
13451
14097
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
13452
14098
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
13453
14099
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
14100
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
14101
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
14102
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
13454
14103
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
14104
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
14105
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
13455
14106
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
13456
14107
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
13457
14108
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -13597,7 +14248,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
13597
14248
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
13598
14249
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
13599
14250
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
14251
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
14252
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
14253
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
13600
14254
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
14255
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
14256
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
13601
14257
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
13602
14258
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
13603
14259
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -13792,7 +14448,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
13792
14448
  OPENAI_API_KEY?: string | undefined;
13793
14449
  OPENAI_BASE_URL?: string | undefined;
13794
14450
  OPENAI_ORGANIZATION?: string | undefined;
14451
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
14452
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
14453
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
13795
14454
  CODEX_API_KEY?: string | undefined;
14455
+ OPENCLAW_CONFIG_PATH?: string | undefined;
14456
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
13796
14457
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
13797
14458
  OPENCLAW_GATEWAY_URL?: string | undefined;
13798
14459
  PALM_API_HOST?: string | undefined;
@@ -14044,7 +14705,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
14044
14705
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
14045
14706
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
14046
14707
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
14708
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
14709
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
14710
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
14047
14711
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
14712
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
14713
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
14048
14714
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
14049
14715
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
14050
14716
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -14171,7 +14837,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
14171
14837
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
14172
14838
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
14173
14839
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
14840
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
14841
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
14842
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
14174
14843
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
14844
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
14845
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
14175
14846
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
14176
14847
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
14177
14848
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -14307,7 +14978,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
14307
14978
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
14308
14979
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
14309
14980
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
14981
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
14982
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
14983
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
14310
14984
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
14985
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
14986
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
14311
14987
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
14312
14988
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
14313
14989
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -14434,7 +15110,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
14434
15110
  OPENAI_API_KEY: z.ZodOptional<z.ZodString>;
14435
15111
  OPENAI_BASE_URL: z.ZodOptional<z.ZodString>;
14436
15112
  OPENAI_ORGANIZATION: z.ZodOptional<z.ZodString>;
15113
+ CLAWDBOT_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
15114
+ CLAWDBOT_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
15115
+ CLAWDBOT_GATEWAY_URL: z.ZodOptional<z.ZodString>;
14437
15116
  CODEX_API_KEY: z.ZodOptional<z.ZodString>;
15117
+ OPENCLAW_CONFIG_PATH: z.ZodOptional<z.ZodString>;
15118
+ OPENCLAW_GATEWAY_PASSWORD: z.ZodOptional<z.ZodString>;
14438
15119
  OPENCLAW_GATEWAY_TOKEN: z.ZodOptional<z.ZodString>;
14439
15120
  OPENCLAW_GATEWAY_URL: z.ZodOptional<z.ZodString>;
14440
15121
  PALM_API_HOST: z.ZodOptional<z.ZodString>;
@@ -14586,7 +15267,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
14586
15267
  OPENAI_API_KEY?: string | undefined;
14587
15268
  OPENAI_BASE_URL?: string | undefined;
14588
15269
  OPENAI_ORGANIZATION?: string | undefined;
15270
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
15271
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
15272
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
14589
15273
  CODEX_API_KEY?: string | undefined;
15274
+ OPENCLAW_CONFIG_PATH?: string | undefined;
15275
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
14590
15276
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
14591
15277
  OPENCLAW_GATEWAY_URL?: string | undefined;
14592
15278
  PALM_API_HOST?: string | undefined;
@@ -14649,7 +15335,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
14649
15335
  prompts?: string[] | undefined;
14650
15336
  providerOutput?: string | Record<string, unknown> | undefined;
14651
15337
  assert?: ({
14652
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
15338
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
14653
15339
  value?: AssertionValue | undefined;
14654
15340
  config?: Record<string, any> | undefined;
14655
15341
  threshold?: number | undefined;
@@ -14665,7 +15351,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
14665
15351
  } | {
14666
15352
  type: "assert-set";
14667
15353
  assert: {
14668
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
15354
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
14669
15355
  value?: AssertionValue | undefined;
14670
15356
  config?: Record<string, any> | undefined;
14671
15357
  threshold?: number | undefined;
@@ -14837,7 +15523,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
14837
15523
  OPENAI_API_KEY?: string | undefined;
14838
15524
  OPENAI_BASE_URL?: string | undefined;
14839
15525
  OPENAI_ORGANIZATION?: string | undefined;
15526
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
15527
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
15528
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
14840
15529
  CODEX_API_KEY?: string | undefined;
15530
+ OPENCLAW_CONFIG_PATH?: string | undefined;
15531
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
14841
15532
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
14842
15533
  OPENCLAW_GATEWAY_URL?: string | undefined;
14843
15534
  PALM_API_HOST?: string | undefined;
@@ -14900,7 +15591,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
14900
15591
  prompts?: string[] | undefined;
14901
15592
  providerOutput?: string | Record<string, unknown> | undefined;
14902
15593
  assert?: ({
14903
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
15594
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
14904
15595
  value?: AssertionValue | undefined;
14905
15596
  config?: Record<string, any> | undefined;
14906
15597
  threshold?: number | undefined;
@@ -14916,7 +15607,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
14916
15607
  } | {
14917
15608
  type: "assert-set";
14918
15609
  assert: {
14919
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
15610
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
14920
15611
  value?: AssertionValue | undefined;
14921
15612
  config?: Record<string, any> | undefined;
14922
15613
  threshold?: number | undefined;
@@ -15087,7 +15778,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
15087
15778
  OPENAI_API_KEY?: string | undefined;
15088
15779
  OPENAI_BASE_URL?: string | undefined;
15089
15780
  OPENAI_ORGANIZATION?: string | undefined;
15781
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
15782
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
15783
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
15090
15784
  CODEX_API_KEY?: string | undefined;
15785
+ OPENCLAW_CONFIG_PATH?: string | undefined;
15786
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
15091
15787
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
15092
15788
  OPENCLAW_GATEWAY_URL?: string | undefined;
15093
15789
  PALM_API_HOST?: string | undefined;
@@ -15150,7 +15846,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
15150
15846
  prompts?: string[] | undefined;
15151
15847
  providerOutput?: string | Record<string, unknown> | undefined;
15152
15848
  assert?: ({
15153
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
15849
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
15154
15850
  value?: AssertionValue | undefined;
15155
15851
  config?: Record<string, any> | undefined;
15156
15852
  threshold?: number | undefined;
@@ -15166,7 +15862,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
15166
15862
  } | {
15167
15863
  type: "assert-set";
15168
15864
  assert: {
15169
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
15865
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
15170
15866
  value?: AssertionValue | undefined;
15171
15867
  config?: Record<string, any> | undefined;
15172
15868
  threshold?: number | undefined;
@@ -15340,7 +16036,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
15340
16036
  OPENAI_API_KEY?: string | undefined;
15341
16037
  OPENAI_BASE_URL?: string | undefined;
15342
16038
  OPENAI_ORGANIZATION?: string | undefined;
16039
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
16040
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
16041
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
15343
16042
  CODEX_API_KEY?: string | undefined;
16043
+ OPENCLAW_CONFIG_PATH?: string | undefined;
16044
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
15344
16045
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
15345
16046
  OPENCLAW_GATEWAY_URL?: string | undefined;
15346
16047
  PALM_API_HOST?: string | undefined;
@@ -15401,7 +16102,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
15401
16102
  } | undefined;
15402
16103
  providerOutput?: string | Record<string, unknown> | undefined;
15403
16104
  assert?: ({
15404
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
16105
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
15405
16106
  value?: AssertionValue | undefined;
15406
16107
  config?: Record<string, any> | undefined;
15407
16108
  threshold?: number | undefined;
@@ -15417,7 +16118,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
15417
16118
  } | {
15418
16119
  type: "assert-set";
15419
16120
  assert: {
15420
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
16121
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
15421
16122
  value?: AssertionValue | undefined;
15422
16123
  config?: Record<string, any> | undefined;
15423
16124
  threshold?: number | undefined;
@@ -15584,7 +16285,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
15584
16285
  OPENAI_API_KEY?: string | undefined;
15585
16286
  OPENAI_BASE_URL?: string | undefined;
15586
16287
  OPENAI_ORGANIZATION?: string | undefined;
16288
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
16289
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
16290
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
15587
16291
  CODEX_API_KEY?: string | undefined;
16292
+ OPENCLAW_CONFIG_PATH?: string | undefined;
16293
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
15588
16294
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
15589
16295
  OPENCLAW_GATEWAY_URL?: string | undefined;
15590
16296
  PALM_API_HOST?: string | undefined;
@@ -15802,7 +16508,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
15802
16508
  OPENAI_API_KEY?: string | undefined;
15803
16509
  OPENAI_BASE_URL?: string | undefined;
15804
16510
  OPENAI_ORGANIZATION?: string | undefined;
16511
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
16512
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
16513
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
15805
16514
  CODEX_API_KEY?: string | undefined;
16515
+ OPENCLAW_CONFIG_PATH?: string | undefined;
16516
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
15806
16517
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
15807
16518
  OPENCLAW_GATEWAY_URL?: string | undefined;
15808
16519
  PALM_API_HOST?: string | undefined;
@@ -15929,7 +16640,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
15929
16640
  OPENAI_API_KEY?: string | undefined;
15930
16641
  OPENAI_BASE_URL?: string | undefined;
15931
16642
  OPENAI_ORGANIZATION?: string | undefined;
16643
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
16644
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
16645
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
15932
16646
  CODEX_API_KEY?: string | undefined;
16647
+ OPENCLAW_CONFIG_PATH?: string | undefined;
16648
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
15933
16649
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
15934
16650
  OPENCLAW_GATEWAY_URL?: string | undefined;
15935
16651
  PALM_API_HOST?: string | undefined;
@@ -16061,7 +16777,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16061
16777
  OPENAI_API_KEY?: string | undefined;
16062
16778
  OPENAI_BASE_URL?: string | undefined;
16063
16779
  OPENAI_ORGANIZATION?: string | undefined;
16780
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
16781
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
16782
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
16064
16783
  CODEX_API_KEY?: string | undefined;
16784
+ OPENCLAW_CONFIG_PATH?: string | undefined;
16785
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
16065
16786
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
16066
16787
  OPENCLAW_GATEWAY_URL?: string | undefined;
16067
16788
  PALM_API_HOST?: string | undefined;
@@ -16188,7 +16909,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16188
16909
  OPENAI_API_KEY?: string | undefined;
16189
16910
  OPENAI_BASE_URL?: string | undefined;
16190
16911
  OPENAI_ORGANIZATION?: string | undefined;
16912
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
16913
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
16914
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
16191
16915
  CODEX_API_KEY?: string | undefined;
16916
+ OPENCLAW_CONFIG_PATH?: string | undefined;
16917
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
16192
16918
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
16193
16919
  OPENCLAW_GATEWAY_URL?: string | undefined;
16194
16920
  PALM_API_HOST?: string | undefined;
@@ -16340,7 +17066,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16340
17066
  OPENAI_API_KEY?: string | undefined;
16341
17067
  OPENAI_BASE_URL?: string | undefined;
16342
17068
  OPENAI_ORGANIZATION?: string | undefined;
17069
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
17070
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
17071
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
16343
17072
  CODEX_API_KEY?: string | undefined;
17073
+ OPENCLAW_CONFIG_PATH?: string | undefined;
17074
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
16344
17075
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
16345
17076
  OPENCLAW_GATEWAY_URL?: string | undefined;
16346
17077
  PALM_API_HOST?: string | undefined;
@@ -16403,7 +17134,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16403
17134
  prompts?: string[] | undefined;
16404
17135
  providerOutput?: string | Record<string, unknown> | undefined;
16405
17136
  assert?: ({
16406
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
17137
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
16407
17138
  value?: AssertionValue | undefined;
16408
17139
  config?: Record<string, any> | undefined;
16409
17140
  threshold?: number | undefined;
@@ -16419,7 +17150,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16419
17150
  } | {
16420
17151
  type: "assert-set";
16421
17152
  assert: {
16422
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
17153
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
16423
17154
  value?: AssertionValue | undefined;
16424
17155
  config?: Record<string, any> | undefined;
16425
17156
  threshold?: number | undefined;
@@ -16591,7 +17322,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16591
17322
  OPENAI_API_KEY?: string | undefined;
16592
17323
  OPENAI_BASE_URL?: string | undefined;
16593
17324
  OPENAI_ORGANIZATION?: string | undefined;
17325
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
17326
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
17327
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
16594
17328
  CODEX_API_KEY?: string | undefined;
17329
+ OPENCLAW_CONFIG_PATH?: string | undefined;
17330
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
16595
17331
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
16596
17332
  OPENCLAW_GATEWAY_URL?: string | undefined;
16597
17333
  PALM_API_HOST?: string | undefined;
@@ -16654,7 +17390,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16654
17390
  prompts?: string[] | undefined;
16655
17391
  providerOutput?: string | Record<string, unknown> | undefined;
16656
17392
  assert?: ({
16657
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
17393
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
16658
17394
  value?: AssertionValue | undefined;
16659
17395
  config?: Record<string, any> | undefined;
16660
17396
  threshold?: number | undefined;
@@ -16670,7 +17406,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16670
17406
  } | {
16671
17407
  type: "assert-set";
16672
17408
  assert: {
16673
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
17409
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
16674
17410
  value?: AssertionValue | undefined;
16675
17411
  config?: Record<string, any> | undefined;
16676
17412
  threshold?: number | undefined;
@@ -16841,7 +17577,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16841
17577
  OPENAI_API_KEY?: string | undefined;
16842
17578
  OPENAI_BASE_URL?: string | undefined;
16843
17579
  OPENAI_ORGANIZATION?: string | undefined;
17580
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
17581
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
17582
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
16844
17583
  CODEX_API_KEY?: string | undefined;
17584
+ OPENCLAW_CONFIG_PATH?: string | undefined;
17585
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
16845
17586
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
16846
17587
  OPENCLAW_GATEWAY_URL?: string | undefined;
16847
17588
  PALM_API_HOST?: string | undefined;
@@ -16904,7 +17645,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16904
17645
  prompts?: string[] | undefined;
16905
17646
  providerOutput?: string | Record<string, unknown> | undefined;
16906
17647
  assert?: ({
16907
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
17648
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
16908
17649
  value?: AssertionValue | undefined;
16909
17650
  config?: Record<string, any> | undefined;
16910
17651
  threshold?: number | undefined;
@@ -16920,7 +17661,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
16920
17661
  } | {
16921
17662
  type: "assert-set";
16922
17663
  assert: {
16923
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
17664
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
16924
17665
  value?: AssertionValue | undefined;
16925
17666
  config?: Record<string, any> | undefined;
16926
17667
  threshold?: number | undefined;
@@ -17094,7 +17835,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
17094
17835
  OPENAI_API_KEY?: string | undefined;
17095
17836
  OPENAI_BASE_URL?: string | undefined;
17096
17837
  OPENAI_ORGANIZATION?: string | undefined;
17838
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
17839
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
17840
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
17097
17841
  CODEX_API_KEY?: string | undefined;
17842
+ OPENCLAW_CONFIG_PATH?: string | undefined;
17843
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
17098
17844
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
17099
17845
  OPENCLAW_GATEWAY_URL?: string | undefined;
17100
17846
  PALM_API_HOST?: string | undefined;
@@ -17155,7 +17901,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
17155
17901
  } | undefined;
17156
17902
  providerOutput?: string | Record<string, unknown> | undefined;
17157
17903
  assert?: ({
17158
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
17904
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
17159
17905
  value?: AssertionValue | undefined;
17160
17906
  config?: Record<string, any> | undefined;
17161
17907
  threshold?: number | undefined;
@@ -17171,7 +17917,7 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
17171
17917
  } | {
17172
17918
  type: "assert-set";
17173
17919
  assert: {
17174
- type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
17920
+ type: `promptfoo:redteam:${string}` | "answer-relevance" | "bleu" | "classifier" | "contains" | "contains-all" | "contains-any" | "contains-html" | "contains-json" | "contains-sql" | "contains-xml" | "context-faithfulness" | "context-recall" | "context-relevance" | "conversation-relevance" | "cost" | "equals" | "factuality" | "finish-reason" | "g-eval" | "gleu" | "guardrails" | "icontains" | "icontains-all" | "icontains-any" | "is-html" | "is-json" | "is-refusal" | "is-sql" | "is-valid-function-call" | "is-valid-openai-function-call" | "is-valid-openai-tools-call" | "is-xml" | "javascript" | "latency" | "levenshtein" | "llm-rubric" | "pi" | "meteor" | "model-graded-closedqa" | "model-graded-factuality" | "moderation" | "perplexity" | "perplexity-score" | "python" | "regex" | "rouge-n" | "ruby" | "similar" | "similar:cosine" | "similar:dot" | "similar:euclidean" | "starts-with" | "tool-call-f1" | "skill-used" | "trajectory:goal-success" | "trajectory:tool-args-match" | "trajectory:step-count" | "trajectory:tool-sequence" | "trajectory:tool-used" | "trace-error-spans" | "trace-span-count" | "trace-span-duration" | "search-rubric" | "webhook" | "word-count" | "not-answer-relevance" | "not-bleu" | "not-classifier" | "not-contains" | "not-contains-all" | "not-contains-any" | "not-contains-html" | "not-contains-json" | "not-contains-sql" | "not-contains-xml" | "not-context-faithfulness" | "not-context-recall" | "not-context-relevance" | "not-conversation-relevance" | "not-cost" | "not-equals" | "not-factuality" | "not-finish-reason" | "not-g-eval" | "not-gleu" | "not-guardrails" | "not-icontains" | "not-icontains-all" | "not-icontains-any" | "not-is-html" | "not-is-json" | "not-is-refusal" | "not-is-sql" | "not-is-valid-function-call" | "not-is-valid-openai-function-call" | "not-is-valid-openai-tools-call" | "not-is-xml" | "not-javascript" | "not-latency" | "not-levenshtein" | "not-llm-rubric" | "not-pi" | "not-meteor" | "not-model-graded-closedqa" | "not-model-graded-factuality" | "not-moderation" | "not-perplexity" | "not-perplexity-score" | "not-python" | "not-regex" | "not-rouge-n" | "not-ruby" | "not-similar" | "not-similar:cosine" | "not-similar:dot" | "not-similar:euclidean" | "not-starts-with" | "not-tool-call-f1" | "not-skill-used" | "not-trajectory:goal-success" | "not-trajectory:tool-args-match" | "not-trajectory:step-count" | "not-trajectory:tool-sequence" | "not-trajectory:tool-used" | "not-trace-error-spans" | "not-trace-span-count" | "not-trace-span-duration" | "not-search-rubric" | "not-webhook" | "not-word-count" | "select-best" | "human" | "max-score";
17175
17921
  value?: AssertionValue | undefined;
17176
17922
  config?: Record<string, any> | undefined;
17177
17923
  threshold?: number | undefined;
@@ -17338,7 +18084,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
17338
18084
  OPENAI_API_KEY?: string | undefined;
17339
18085
  OPENAI_BASE_URL?: string | undefined;
17340
18086
  OPENAI_ORGANIZATION?: string | undefined;
18087
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
18088
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
18089
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
17341
18090
  CODEX_API_KEY?: string | undefined;
18091
+ OPENCLAW_CONFIG_PATH?: string | undefined;
18092
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
17342
18093
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
17343
18094
  OPENCLAW_GATEWAY_URL?: string | undefined;
17344
18095
  PALM_API_HOST?: string | undefined;
@@ -17556,7 +18307,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
17556
18307
  OPENAI_API_KEY?: string | undefined;
17557
18308
  OPENAI_BASE_URL?: string | undefined;
17558
18309
  OPENAI_ORGANIZATION?: string | undefined;
18310
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
18311
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
18312
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
17559
18313
  CODEX_API_KEY?: string | undefined;
18314
+ OPENCLAW_CONFIG_PATH?: string | undefined;
18315
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
17560
18316
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
17561
18317
  OPENCLAW_GATEWAY_URL?: string | undefined;
17562
18318
  PALM_API_HOST?: string | undefined;
@@ -17683,7 +18439,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
17683
18439
  OPENAI_API_KEY?: string | undefined;
17684
18440
  OPENAI_BASE_URL?: string | undefined;
17685
18441
  OPENAI_ORGANIZATION?: string | undefined;
18442
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
18443
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
18444
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
17686
18445
  CODEX_API_KEY?: string | undefined;
18446
+ OPENCLAW_CONFIG_PATH?: string | undefined;
18447
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
17687
18448
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
17688
18449
  OPENCLAW_GATEWAY_URL?: string | undefined;
17689
18450
  PALM_API_HOST?: string | undefined;
@@ -17815,7 +18576,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
17815
18576
  OPENAI_API_KEY?: string | undefined;
17816
18577
  OPENAI_BASE_URL?: string | undefined;
17817
18578
  OPENAI_ORGANIZATION?: string | undefined;
18579
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
18580
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
18581
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
17818
18582
  CODEX_API_KEY?: string | undefined;
18583
+ OPENCLAW_CONFIG_PATH?: string | undefined;
18584
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
17819
18585
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
17820
18586
  OPENCLAW_GATEWAY_URL?: string | undefined;
17821
18587
  PALM_API_HOST?: string | undefined;
@@ -17942,7 +18708,12 @@ declare const UnifiedConfigSchema: z.ZodPipe<z.ZodObject<{
17942
18708
  OPENAI_API_KEY?: string | undefined;
17943
18709
  OPENAI_BASE_URL?: string | undefined;
17944
18710
  OPENAI_ORGANIZATION?: string | undefined;
18711
+ CLAWDBOT_GATEWAY_PASSWORD?: string | undefined;
18712
+ CLAWDBOT_GATEWAY_TOKEN?: string | undefined;
18713
+ CLAWDBOT_GATEWAY_URL?: string | undefined;
17945
18714
  CODEX_API_KEY?: string | undefined;
18715
+ OPENCLAW_CONFIG_PATH?: string | undefined;
18716
+ OPENCLAW_GATEWAY_PASSWORD?: string | undefined;
17946
18717
  OPENCLAW_GATEWAY_TOKEN?: string | undefined;
17947
18718
  OPENCLAW_GATEWAY_URL?: string | undefined;
17948
18719
  PALM_API_HOST?: string | undefined;
@@ -18158,7 +18929,8 @@ declare function runAssertion({
18158
18929
  vars,
18159
18930
  latencyMs,
18160
18931
  providerResponse,
18161
- traceId
18932
+ traceId,
18933
+ traceData
18162
18934
  }: {
18163
18935
  prompt?: string;
18164
18936
  provider?: ApiProvider;
@@ -18169,6 +18941,7 @@ declare function runAssertion({
18169
18941
  latencyMs?: number;
18170
18942
  assertIndex?: number;
18171
18943
  traceId?: string;
18944
+ traceData?: TraceData | null;
18172
18945
  }): Promise<GradingResult>;
18173
18946
  declare function runAssertions({
18174
18947
  assertScoringFunction,
@@ -18834,5 +19607,5 @@ declare const _default$1: {
18834
19607
  };
18835
19608
  };
18836
19609
  //#endregion
18837
- export { type AfterAllExtensionHookContext, type AfterEachExtensionHookContext, AgentCancelledPayload, AgentErrorPayload, ApiClassificationProvider, ApiEmbeddingProvider, ApiModerationProvider, ApiProvider, ApiSimilarityProvider, Assertion, AssertionOrSet, AssertionOrSetSchema, AssertionParams, AssertionSchema, AssertionSet, AssertionSetSchema, AssertionType, AssertionTypeSchema, AssertionValue, AssertionValueFunction, AssertionValueFunctionContext, AssertionValueFunctionResult, AtomicTestCase, AtomicTestCaseSchema, AudioGradingConfig, BaseAssertionTypes, BaseAssertionTypesSchema, BaseRedteamMetadata, BaseTokenUsage, BaseTokenUsageSchema, type BeforeAllExtensionHookContext, type BeforeEachExtensionHookContext, CallApiContextParams, CallApiFunction, CallApiOptionsParams, ChatMessage, CommandLineOptions, CommandLineOptionsSchema, CompletedPrompt, CompletedPromptSchema, CompletionTokenDetails, CompletionTokenDetailsSchema, ConversationMessage, ConversationMessageSchema, CsvRow, DefaultProviders, DerivedMetric, DerivedMetricSchema, type EnvOverrides, EvalConversations, EvalRegisters, EvalResultsFilterMode, EvalSummary, EvalTableDTO, EvalWithMetadata, EvaluateOptions, EvaluateOptionsSchema, EvaluateResult, EvaluateStats, EvaluateSummaryV2, EvaluateSummaryV3, EvaluateTable, EvaluateTableOutput, EvaluateTableRow, EvaluateTestSuite, EvaluateTestSuiteWithEvaluateOptions, type ExtensionHookContextMap, FailedPluginInfo, FilePath, GradingConfig, GradingConfigSchema, GradingResult, GuardrailResponse, ImageOutput, InferSchema, type Inputs, InputsSchema, Intent, Job, LoadApiProviderContext, ModerationFlag, Modifier, NotPrefixedAssertionTypesSchema, NunjucksFilterMap, OutputConfig, OutputConfigSchema, OutputFile, OutputFileExtension, OutputMetadata, PartialGenerationError, PluginActionParams, PluginConfig, PluginConfigSchema, PoliciesById, Policy, PolicyObject, PolicyObjectSchema, Prompt, PromptConfig, PromptContent, PromptFunction, PromptFunctionContext, PromptFunctionResult, PromptMetrics, PromptWithMetadata, ProviderClassificationResponse, ProviderEmbeddingResponse, ProviderFunction, ProviderId, ProviderLabel, ProviderModerationResponse, ProviderOptions, ProviderOptionsMap, ProviderResponse, ProviderSimilarityResponse, ProviderTestResponse, ProviderType, ProviderTypeMap, ProvidersSchema, RateLimitRegistryRef, RedteamAssertionTypes, RedteamCliGenerateOptions, RedteamContext, RedteamFileConfig, RedteamGenerateOptions, RedteamHistoryEntry, RedteamMediaData, RedteamObjectConfig, RedteamPlugin, RedteamPluginObject, RedteamRunOptions, RedteamStrategy, RedteamStrategyObject, ResultFailureReason, ResultLightweight, ResultLightweightWithLabel, ResultSuggestion, ResultsFile, RunEvalOptions, SavedRedteamConfig, Scenario, ScenarioSchema, ScoringFunction, ServerPromptWithMetadata, SharedResults, SpecialAssertionTypes, SpecialAssertionTypesSchema, StrategyConfig, StrategyConfigSchema, SynthesizeOptions, TestCase, TestCaseSchema, TestCaseWithPlugin, TestCaseWithVarsFile, TestCaseWithVarsFileSchema, TestCasesWithMetadata, TestCasesWithMetadataPrompt, TestCasesWithMetadataPromptSchema, TestCasesWithMetadataSchema, TestGeneratorConfig, TestGeneratorConfigSchema, TestSuite, TestSuiteConfig, TestSuiteConfigSchema, TestSuiteSchema, type TokenUsage, TraceData, TraceSpan, TracingConfig, UnifiedConfig, UnifiedConfigSchema, VarMapping, type VarValue, Vars, VarsSchema, ZodLikeSchema, _default as assertions, cache_d_exports as cache, _default$1 as default, evaluate, generateTable, guardrails, isApiProvider, isGradingResult, isProviderOptions, isResultFailureReason, loadApiProvider, redteam };
19610
+ export { type AfterAllExtensionHookContext, type AfterEachExtensionHookContext, AgentCancelledPayload, AgentErrorPayload, ApiClassificationProvider, ApiEmbeddingProvider, ApiModerationProvider, ApiProvider, ApiSimilarityProvider, Assertion, AssertionOrSet, AssertionOrSetSchema, AssertionParams, AssertionSchema, AssertionSet, AssertionSetSchema, AssertionType, AssertionTypeSchema, AssertionValue, AssertionValueFunction, AssertionValueFunctionContext, AssertionValueFunctionResult, AtomicTestCase, AtomicTestCaseSchema, AudioGradingConfig, BaseAssertionTypes, BaseAssertionTypesSchema, BaseRedteamMetadata, BaseTokenUsage, BaseTokenUsageSchema, type BeforeAllExtensionHookContext, type BeforeEachExtensionHookContext, CallApiContextParams, CallApiFunction, CallApiOptionsParams, ChatMessage, CommandLineOptions, CommandLineOptionsSchema, CompletedPrompt, CompletedPromptSchema, CompletionTokenDetails, CompletionTokenDetailsSchema, ConversationMessage, ConversationMessageSchema, CsvRow, DefaultProviders, DerivedMetric, DerivedMetricSchema, type EnvOverrides, EvalConversations, EvalRegisters, EvalResultsFilterMode, EvalSummary, EvalTableDTO, EvalWithMetadata, EvaluateOptions, EvaluateOptionsSchema, EvaluateResult, EvaluateStats, EvaluateSummaryV2, EvaluateSummaryV3, EvaluateTable, EvaluateTableOutput, EvaluateTableRow, EvaluateTestSuite, EvaluateTestSuiteWithEvaluateOptions, type ExtensionHookContextMap, FailedPluginInfo, FilePath, GradingConfig, GradingConfigSchema, GradingResult, GuardrailResponse, ImageOutput, InferSchema, type Inputs, InputsSchema, Intent, Job, LoadApiProviderContext, ModerationFlag, Modifier, NotPrefixedAssertionTypesSchema, NunjucksFilterMap, OutputConfig, OutputConfigSchema, OutputFile, OutputFileExtension, OutputMetadata, PartialGenerationError, PluginActionParams, PluginConfig, PluginConfigSchema, PoliciesById, Policy, PolicyObject, PolicyObjectSchema, Prompt, PromptConfig, PromptContent, PromptFunction, PromptFunctionContext, PromptFunctionResult, PromptMetrics, PromptWithMetadata, ProviderClassificationResponse, ProviderEmbeddingResponse, ProviderFunction, ProviderId, ProviderLabel, ProviderModerationResponse, ProviderOptions, ProviderOptionsMap, ProviderResponse, ProviderSimilarityResponse, ProviderTestResponse, ProviderType, ProviderTypeMap, ProvidersSchema, RateLimitRegistryRef, RedteamAssertionTypes, RedteamCliGenerateOptions, RedteamContext, RedteamFileConfig, RedteamGenerateOptions, RedteamHistoryEntry, RedteamMediaData, RedteamObjectConfig, RedteamPlugin, RedteamPluginObject, RedteamRunOptions, RedteamStrategy, RedteamStrategyObject, ResultFailureReason, ResultLightweight, ResultLightweightWithLabel, ResultSuggestion, ResultsFile, RunEvalOptions, SavedRedteamConfig, Scenario, ScenarioSchema, ScoringFunction, ServerPromptWithMetadata, SharedResults, SkillCallEntry, SpecialAssertionTypes, SpecialAssertionTypesSchema, StrategyConfig, StrategyConfigSchema, SynthesizeOptions, TestCase, TestCaseSchema, TestCaseWithPlugin, TestCaseWithVarsFile, TestCaseWithVarsFileSchema, TestCasesWithMetadata, TestCasesWithMetadataPrompt, TestCasesWithMetadataPromptSchema, TestCasesWithMetadataSchema, TestGeneratorConfig, TestGeneratorConfigSchema, TestSuite, TestSuiteConfig, TestSuiteConfigSchema, TestSuiteSchema, type TokenUsage, TraceData, TraceSpan, TracingConfig, UnifiedConfig, UnifiedConfigSchema, VarMapping, type VarValue, Vars, VarsSchema, ZodLikeSchema, _default as assertions, cache_d_exports as cache, _default$1 as default, evaluate, generateTable, guardrails, isApiProvider, isGradingResult, isProviderOptions, isResultFailureReason, loadApiProvider, redteam };
18838
19611
  //# sourceMappingURL=index.d.ts.map