@tangle-network/agent-eval 0.145.22 → 0.147.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (153) hide show
  1. package/CHANGELOG.md +34 -0
  2. package/dist/analyst/index.d.ts +13 -13
  3. package/dist/analyst/index.js +6 -6
  4. package/dist/{backend-integrity-HVmWyOaX.d.ts → backend-integrity-Bz8nSrRE.d.ts} +2 -2
  5. package/dist/{backend-integrity-HVmWyOaX.d.ts.map → backend-integrity-Bz8nSrRE.d.ts.map} +1 -1
  6. package/dist/{benchmark-DTQ1RApl.d.ts → benchmark-Cn7ZMFNW.d.ts} +3 -3
  7. package/dist/{benchmark-DTQ1RApl.d.ts.map → benchmark-Cn7ZMFNW.d.ts.map} +1 -1
  8. package/dist/{benchmark-command-CDB_yKqS.js → benchmark-command-BTiysiYU.js} +11 -11
  9. package/dist/{benchmark-command-CDB_yKqS.js.map → benchmark-command-BTiysiYU.js.map} +1 -1
  10. package/dist/benchmarks/index.d.ts +4 -4
  11. package/dist/benchmarks/index.js +3 -3
  12. package/dist/campaign/index.d.ts +8 -8
  13. package/dist/campaign/index.js +5 -5
  14. package/dist/{campaign-CbQXcARJ.js → campaign-DCDdhuv2.js} +7 -7
  15. package/dist/{campaign-CbQXcARJ.js.map → campaign-DCDdhuv2.js.map} +1 -1
  16. package/dist/{capture-fetch-BZiO2cEH.d.ts → capture-fetch-BAx3Gntt.d.ts} +2 -2
  17. package/dist/{capture-fetch-BZiO2cEH.d.ts.map → capture-fetch-BAx3Gntt.d.ts.map} +1 -1
  18. package/dist/{chat-client-BJmwfnjN.js → chat-client-BX8Wh3fn.js} +3 -3
  19. package/dist/{chat-client-BJmwfnjN.js.map → chat-client-BX8Wh3fn.js.map} +1 -1
  20. package/dist/cli.js +2 -2
  21. package/dist/{client-YDkU5QST.d.ts → client-BV7icIfy.d.ts} +2 -2
  22. package/dist/{client-YDkU5QST.d.ts.map → client-BV7icIfy.d.ts.map} +1 -1
  23. package/dist/contract/index.d.ts +10 -10
  24. package/dist/contract/index.js +9 -9
  25. package/dist/{cost-ledger-BSe92yAV.js → cost-ledger-B1qx30B4.js} +2 -2
  26. package/dist/{cost-ledger-BSe92yAV.js.map → cost-ledger-B1qx30B4.js.map} +1 -1
  27. package/dist/{default-registry-DONui3mA.d.ts → default-registry-BUtRMRKk.d.ts} +6 -6
  28. package/dist/{default-registry-DONui3mA.d.ts.map → default-registry-BUtRMRKk.d.ts.map} +1 -1
  29. package/dist/{define-agent-eval-BGxmy4W6.js → define-agent-eval-BqWFz3sK.js} +3 -3
  30. package/dist/{define-agent-eval-BGxmy4W6.js.map → define-agent-eval-BqWFz3sK.js.map} +1 -1
  31. package/dist/{define-agent-eval-DrpgTVPp.d.ts → define-agent-eval-D52ClbX2.d.ts} +5 -5
  32. package/dist/{define-agent-eval-DrpgTVPp.d.ts.map → define-agent-eval-D52ClbX2.d.ts.map} +1 -1
  33. package/dist/{dspy-rlm-engine-B2nst4NR.js → dspy-rlm-engine-BmtOl_kP.js} +4 -4
  34. package/dist/{dspy-rlm-engine-B2nst4NR.js.map → dspy-rlm-engine-BmtOl_kP.js.map} +1 -1
  35. package/dist/{engine-Bc3seRPe.d.ts → engine-CIy18RTX.d.ts} +4 -4
  36. package/dist/{engine-Bc3seRPe.d.ts.map → engine-CIy18RTX.d.ts.map} +1 -1
  37. package/dist/{eval-campaign-DeGLwACc.js → eval-campaign-bmZ6NIIP.js} +2 -2
  38. package/dist/{eval-campaign-DeGLwACc.js.map → eval-campaign-bmZ6NIIP.js.map} +1 -1
  39. package/dist/{exact-types-CYFbvA5U.d.ts → exact-types-rdKFzEnK.d.ts} +2 -2
  40. package/dist/{exact-types-CYFbvA5U.d.ts.map → exact-types-rdKFzEnK.d.ts.map} +1 -1
  41. package/dist/experiment/index.d.ts +5 -5
  42. package/dist/experiment/index.js +3 -3
  43. package/dist/{experiment-tracker-DWHZBAYL.d.ts → experiment-tracker-D9VHwRm-.d.ts} +6 -24
  44. package/dist/experiment-tracker-D9VHwRm-.d.ts.map +1 -0
  45. package/dist/{experiment-tracker-C29gXM4B.js → experiment-tracker-Ym6rEQT1.js} +15 -4
  46. package/dist/experiment-tracker-Ym6rEQT1.js.map +1 -0
  47. package/dist/{external-optimizer-contracts-CKEY98bK.d.ts → external-optimizer-contracts-CSjDLLmr.d.ts} +2 -2
  48. package/dist/{external-optimizer-contracts-CKEY98bK.d.ts.map → external-optimizer-contracts-CSjDLLmr.d.ts.map} +1 -1
  49. package/dist/{external-optimizer-process-DYNizz8W.js → external-optimizer-process-C4AyCcQd.js} +2 -2
  50. package/dist/{external-optimizer-process-DYNizz8W.js.map → external-optimizer-process-C4AyCcQd.js.map} +1 -1
  51. package/dist/{external-optimizer-subprocess-r6dP9DAj.js → external-optimizer-subprocess-DU1KM8yV.js} +3 -3
  52. package/dist/{external-optimizer-subprocess-r6dP9DAj.js.map → external-optimizer-subprocess-DU1KM8yV.js.map} +1 -1
  53. package/dist/{feedback-trajectory-Cg9EDxKB.d.ts → feedback-trajectory-niGeR6uJ.d.ts} +3 -3
  54. package/dist/{feedback-trajectory-Cg9EDxKB.d.ts.map → feedback-trajectory-niGeR6uJ.d.ts.map} +1 -1
  55. package/dist/fuzz.js +1 -1
  56. package/dist/hosted/index.d.ts +2 -2
  57. package/dist/{index-DGrcBMbC.d.ts → index-BjBjxiVv.d.ts} +9 -9
  58. package/dist/{index-DGrcBMbC.d.ts.map → index-BjBjxiVv.d.ts.map} +1 -1
  59. package/dist/index-CKblBxtr.d.ts +1 -0
  60. package/dist/{index-CevPFhDT.d.ts → index-gtjtfLEJ.d.ts} +7 -7
  61. package/dist/{index-CevPFhDT.d.ts.map → index-gtjtfLEJ.d.ts.map} +1 -1
  62. package/dist/index.d.ts +21 -51
  63. package/dist/index.d.ts.map +1 -1
  64. package/dist/index.js +16 -125
  65. package/dist/index.js.map +1 -1
  66. package/dist/{integrity-DDoeWibF.d.ts → integrity-CGfpTE5-.d.ts} +2 -2
  67. package/dist/{integrity-DDoeWibF.d.ts.map → integrity-CGfpTE5-.d.ts.map} +1 -1
  68. package/dist/{kind-factory-CPmSd58s.js → kind-factory-DmAa0h3K.js} +7 -7
  69. package/dist/kind-factory-DmAa0h3K.js.map +1 -0
  70. package/dist/{llm-client-_UE4fU7Z.js → llm-client-BMuxYoZy.js} +2 -2
  71. package/dist/{llm-client-_UE4fU7Z.js.map → llm-client-BMuxYoZy.js.map} +1 -1
  72. package/dist/{llm-judge-B7ffRF8w.js → llm-judge-DrzsVS5k.js} +4 -4
  73. package/dist/{llm-judge-B7ffRF8w.js.map → llm-judge-DrzsVS5k.js.map} +1 -1
  74. package/dist/matrix-C-2Qx1Zr.d.ts +383 -0
  75. package/dist/matrix-C-2Qx1Zr.d.ts.map +1 -0
  76. package/dist/meta-eval/index.d.ts +1 -1
  77. package/dist/{metrics-Cl0L1KUy.js → metrics-Qv-cpptD.js} +15 -8
  78. package/dist/metrics-Qv-cpptD.js.map +1 -0
  79. package/dist/multishot/golden/index.d.ts +230 -0
  80. package/dist/multishot/golden/index.d.ts.map +1 -0
  81. package/dist/multishot/golden/index.js +7686 -0
  82. package/dist/multishot/golden/index.js.map +1 -0
  83. package/dist/multishot/index.d.ts +2 -380
  84. package/dist/multishot/index.d.ts.map +1 -1
  85. package/dist/multishot/index.js +2 -106
  86. package/dist/multishot/index.js.map +1 -1
  87. package/dist/openapi.json +1 -1
  88. package/dist/{produced-state-D3k45S9a.js → produced-state-Dtx60bUQ.js} +4 -4
  89. package/dist/{produced-state-D3k45S9a.js.map → produced-state-Dtx60bUQ.js.map} +1 -1
  90. package/dist/{promotion-policy-BKi5OK5l.d.ts → promotion-policy-CewdjfCJ.d.ts} +2 -2
  91. package/dist/{promotion-policy-BKi5OK5l.d.ts.map → promotion-policy-CewdjfCJ.d.ts.map} +1 -1
  92. package/dist/{provenance-Bb7Zmyj2.d.ts → provenance-LrOEOHQb.d.ts} +6 -6
  93. package/dist/{provenance-Bb7Zmyj2.d.ts.map → provenance-LrOEOHQb.d.ts.map} +1 -1
  94. package/dist/{registry-CZgEmVXb.d.ts → registry-CaqJIDdN.d.ts} +4 -4
  95. package/dist/{registry-CZgEmVXb.d.ts.map → registry-CaqJIDdN.d.ts.map} +1 -1
  96. package/dist/reporting.d.ts +1 -1
  97. package/dist/reporting.js +1 -1
  98. package/dist/{researcher-CC305Ed-.d.ts → researcher-BySlpla_.d.ts} +3 -3
  99. package/dist/{researcher-CC305Ed-.d.ts.map → researcher-BySlpla_.d.ts.map} +1 -1
  100. package/dist/rl.d.ts +3 -3
  101. package/dist/rl.js +2 -2
  102. package/dist/{semantic-concept-judge-BsY2Q0Oq.js → semantic-concept-judge-BI7Rrl5-.js} +3 -3
  103. package/dist/{semantic-concept-judge-BsY2Q0Oq.js.map → semantic-concept-judge-BI7Rrl5-.js.map} +1 -1
  104. package/dist/{sequential-CYwq6Ff_.d.ts → sequential-BhsrMupG.d.ts} +31 -2
  105. package/dist/sequential-BhsrMupG.d.ts.map +1 -0
  106. package/dist/{sequential-Br0mAPHA.js → sequential-CzK5DarL.js} +48 -6
  107. package/dist/sequential-CzK5DarL.js.map +1 -0
  108. package/dist/series-convergence-CjO2QdRW.js.map +1 -1
  109. package/dist/{series-convergence-C9G-GNYK.d.ts → series-convergence-D2fsoJ2w.d.ts} +4 -7
  110. package/dist/{series-convergence-C9G-GNYK.d.ts.map → series-convergence-D2fsoJ2w.d.ts.map} +1 -1
  111. package/dist/{server-cQXve8i4.js → server-D9wQclzG.js} +3 -3
  112. package/dist/{server-cQXve8i4.js.map → server-D9wQclzG.js.map} +1 -1
  113. package/dist/{skillopt-optimization-method-C9MrMgW9.d.ts → skillopt-optimization-method-BoC1Qccx.d.ts} +5 -5
  114. package/dist/{skillopt-optimization-method-C9MrMgW9.d.ts.map → skillopt-optimization-method-BoC1Qccx.d.ts.map} +1 -1
  115. package/dist/{skillopt-optimization-method-Cv3K4DaS.js → skillopt-optimization-method-C_UrqZs2.js} +4 -4
  116. package/dist/{skillopt-optimization-method-Cv3K4DaS.js.map → skillopt-optimization-method-C_UrqZs2.js.map} +1 -1
  117. package/dist/{statistical-heldout-fvxtHnKQ.d.ts → statistical-heldout-uxSpFEjm.d.ts} +2 -2
  118. package/dist/{statistical-heldout-fvxtHnKQ.d.ts.map → statistical-heldout-uxSpFEjm.d.ts.map} +1 -1
  119. package/dist/{store-otlp-CsptLYpN.js → store-otlp-CDYWW_8N.js} +2 -2
  120. package/dist/{store-otlp-CsptLYpN.js.map → store-otlp-CDYWW_8N.js.map} +1 -1
  121. package/dist/{store-tool-spans-Cq9mFd-q.js → store-tool-spans-CykkbOlv.js} +3 -3
  122. package/dist/{store-tool-spans-Cq9mFd-q.js.map → store-tool-spans-CykkbOlv.js.map} +1 -1
  123. package/dist/{store-tool-spans-CV0hRsUp.d.ts → store-tool-spans-D5FhM0_A.d.ts} +4 -4
  124. package/dist/{store-tool-spans-CV0hRsUp.d.ts.map → store-tool-spans-D5FhM0_A.d.ts.map} +1 -1
  125. package/dist/{task-failure-attributes-CpQ4y5RD.js → task-failure-attributes--ZTP3tYO.js} +2 -2
  126. package/dist/{task-failure-attributes-CpQ4y5RD.js.map → task-failure-attributes--ZTP3tYO.js.map} +1 -1
  127. package/dist/{tool-groups-CDDXNKhd.d.ts → tool-groups-zpufabP8.d.ts} +3 -3
  128. package/dist/tool-groups-zpufabP8.d.ts.map +1 -0
  129. package/dist/trace-repair/index.d.ts +1 -1
  130. package/dist/traces.d.ts +7 -7
  131. package/dist/traces.js +4 -4
  132. package/dist/types-B3jzCp0p.js +108 -0
  133. package/dist/types-B3jzCp0p.js.map +1 -0
  134. package/dist/{types-oAsa-fs4.d.ts → types-BohHewKK.d.ts} +2 -2
  135. package/dist/{types-oAsa-fs4.d.ts.map → types-BohHewKK.d.ts.map} +1 -1
  136. package/dist/{types-DABZgDGV.d.ts → types-C1Bmeb8X.d.ts} +3 -3
  137. package/dist/{types-DABZgDGV.d.ts.map → types-C1Bmeb8X.d.ts.map} +1 -1
  138. package/dist/{types-CWcPguwd.d.ts → types-ColZlZtc.d.ts} +2 -2
  139. package/dist/{types-CWcPguwd.d.ts.map → types-ColZlZtc.d.ts.map} +1 -1
  140. package/dist/{types-XBRdmbj8.d.ts → types-DOhGq4S1.d.ts} +1 -7
  141. package/dist/{types-XBRdmbj8.d.ts.map → types-DOhGq4S1.d.ts.map} +1 -1
  142. package/dist/wire/index.d.ts +2 -2
  143. package/dist/wire/index.js +1 -1
  144. package/docs/multishot-golden-records.md +102 -0
  145. package/package.json +7 -2
  146. package/dist/experiment-tracker-C29gXM4B.js.map +0 -1
  147. package/dist/experiment-tracker-DWHZBAYL.d.ts.map +0 -1
  148. package/dist/index-BgHYx1QG.d.ts +0 -1
  149. package/dist/kind-factory-CPmSd58s.js.map +0 -1
  150. package/dist/metrics-Cl0L1KUy.js.map +0 -1
  151. package/dist/sequential-Br0mAPHA.js.map +0 -1
  152. package/dist/sequential-CYwq6Ff_.d.ts.map +0 -1
  153. package/dist/tool-groups-CDDXNKhd.d.ts.map +0 -1
@@ -1,5 +1,5 @@
1
1
  import { u as RunTokenUsage } from "./run-record-DVV82Gwh.js";
2
- import { dt as ProviderRedactor, mt as RawProviderSink } from "./types-XBRdmbj8.js";
2
+ import { dt as ProviderRedactor, mt as RawProviderSink } from "./types-DOhGq4S1.js";
3
3
  //#region src/trace/extract-usage.d.ts
4
4
  type ExtractedUsage = RunTokenUsage;
5
5
  type SseUsageMode = 'cumulative' | 'delta';
@@ -65,4 +65,4 @@ interface CaptureFetchOptions {
65
65
  declare function captureFetchToRawSink(fetch: typeof globalThis.fetch, sink: RawProviderSink, ctx: CaptureFetchContext, opts?: CaptureFetchOptions): typeof globalThis.fetch;
66
66
  //#endregion
67
67
  export { ExtractedUsage as a, extractUsageFromResponse as c, ExtractUsageFromSseOptions as i, extractUsageFromSse as l, CaptureFetchOptions as n, SseUsageMode as o, captureFetchToRawSink as r, extractUsage as s, CaptureFetchContext as t };
68
- //# sourceMappingURL=capture-fetch-BZiO2cEH.d.ts.map
68
+ //# sourceMappingURL=capture-fetch-BAx3Gntt.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"capture-fetch-BZiO2cEH.d.ts","names":[],"sources":["../src/trace/extract-usage.ts","../src/trace/capture-fetch.ts"],"mappings":";;;KAgBY,iBAAiB;KACjB;UAEK;;EAEf,OAAO;;;;;;;iBAuDO,aAAa,gBAAgB;;;;;;iBA6C7B,oBACd,cACA,UAAS,6BACR;;;;;;;iBA4DmB,yBACpB,UAAU,UACV,aAAa,6BACZ,QAAQ;;;UCrKM;;;EAGf;;EAEA;;;EAGA;;EAEA;;EAEA;;UAGe;;EAEf,WAAW;;;EAGX;;;EAGA;;;;;;;;EAQA,WAAW,OAAO,gBAAgB,KAAK;;EAEvC,eAAe;;iBAkDD,sBACd,cAAc,WAAW,OACzB,MAAM,iBACN,KAAK,qBACL,OAAM,6BACE,WAAW"}
1
+ {"version":3,"file":"capture-fetch-BAx3Gntt.d.ts","names":[],"sources":["../src/trace/extract-usage.ts","../src/trace/capture-fetch.ts"],"mappings":";;;KAgBY,iBAAiB;KACjB;UAEK;;EAEf,OAAO;;;;;;;iBAuDO,aAAa,gBAAgB;;;;;;iBA6C7B,oBACd,cACA,UAAS,6BACR;;;;;;;iBA4DmB,yBACpB,UAAU,UACV,aAAa,6BACZ,QAAQ;;;UCrKM;;;EAGf;;EAEA;;;EAGA;;EAEA;;EAEA;;UAGe;;EAEf,WAAW;;;EAGX;;;EAGA;;;;;;;;EAQA,WAAW,OAAO,gBAAgB,KAAK;;EAEvC,eAAe;;iBAkDD,sBACd,cAAc,WAAW,OACzB,MAAM,iBACN,KAAK,qBACL,OAAM,6BACE,WAAW"}
@@ -1,9 +1,9 @@
1
1
  import { i as hashCanonical, r as canonicalString } from "./canonical-D-XsTQ6_.js";
2
2
  import { a as combineAbortSignals } from "./run-score-lDzV0X8j.js";
3
3
  import { c as deepFreezeCanonicalJson, i as assertValidAnalystUsageReceipt, n as makeFinding, s as validateUsageSettlementTimeout } from "./types-BI4fT3HN.js";
4
- import { R as findingSubjectGrammarPromptFor, X as spanEpochMillis, nt as snapshotExactExecutionComponentIdentity, rt as snapshotExactExecutionPlan, t as createTraceAnalyst } from "./kind-factory-CPmSd58s.js";
4
+ import { R as findingSubjectGrammarPromptFor, X as spanEpochMillis, nt as snapshotExactExecutionComponentIdentity, rt as snapshotExactExecutionPlan, t as createTraceAnalyst } from "./kind-factory-DmAa0h3K.js";
5
5
  import { LLM_CONTEXT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_OUTPUT_TOKEN_ATTR_KEYS, TOOL_NAME_ATTR_KEYS } from "./trace-attributes.js";
6
- import { n as LlmClient } from "./llm-client-_UE4fU7Z.js";
6
+ import { n as LlmClient } from "./llm-client-BMuxYoZy.js";
7
7
  import { t as executionTrackByLane } from "./execution-tracks-CpgFPpS5.js";
8
8
  import { t as analyzeSupervisorRunIntegrity } from "./integrity-DysDBWDu.js";
9
9
  import { randomUUID } from "node:crypto";
@@ -2449,4 +2449,4 @@ function resolveModel(req, defaultModel) {
2449
2449
  //#endregion
2450
2450
  export { validateAnalystReviewDecisions as S, analystRunDigest as _, assertExactRegistryRunOpts as a, readAnalystReview as b, KNOWLEDGE_GAP_KIND_SPEC as c, CONTROL_INTEGRITY_ANALYST as d, ControlIntegrityAnalyst as f, analystFindingDigest as g, deriveEfficiencyFindings as h, ExactAnalystRunExecutionError as i, IMPROVEMENT_KIND_SPEC as l, behavioralAnalyst as m, buildDefaultAnalystRegistry as n, DEFAULT_TRACE_ANALYST_KINDS as o, emitControlIntegrityFindings as p, AnalystRegistry as r, KNOWLEDGE_POISONING_KIND_SPEC as s, createChatClient as t, FAILURE_MODE_KIND_SPEC as u, assertUniqueFindingIds as v, snapshotAnalystRun as x, completedAnalystReviewQuality as y };
2451
2451
 
2452
- //# sourceMappingURL=chat-client-BJmwfnjN.js.map
2452
+ //# sourceMappingURL=chat-client-BX8Wh3fn.js.map