@tangle-network/agent-eval 0.184.0 → 0.186.0

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 (133) hide show
  1. package/CHANGELOG.md +20 -0
  2. package/dist/adapters/http.d.ts +1 -1
  3. package/dist/analyst/index.d.ts +6 -331
  4. package/dist/analyst/index.d.ts.map +1 -1
  5. package/dist/analyst/index.js +7 -6
  6. package/dist/analyst/index.js.map +1 -1
  7. package/dist/{define-agent-eval-RTekxhUE.js → analyze-runs-DCY7LMt4.js} +4 -811
  8. package/dist/analyze-runs-DCY7LMt4.js.map +1 -0
  9. package/dist/analyze-runs-DapYZz02.d.ts +75 -0
  10. package/dist/analyze-runs-DapYZz02.d.ts.map +1 -0
  11. package/dist/{benchmark-command-DlSePljH.js → benchmark-command-tBgs-_AM.js} +9 -47
  12. package/dist/benchmark-command-tBgs-_AM.js.map +1 -0
  13. package/dist/benchmarks/index.d.ts +1 -1
  14. package/dist/benchmarks/index.js +2 -2
  15. package/dist/campaign/index.d.ts +2 -2
  16. package/dist/campaign/index.js +7 -6
  17. package/dist/{campaign-D6puiGFO.js → campaign-DlxdTOvw.js} +8 -7
  18. package/dist/{campaign-D6puiGFO.js.map → campaign-DlxdTOvw.js.map} +1 -1
  19. package/dist/{campaign-evidence-aWEmfROB.js → campaign-evidence-IF1p0oDa.js} +4 -255
  20. package/dist/campaign-evidence-IF1p0oDa.js.map +1 -0
  21. package/dist/cli.js +1 -1
  22. package/dist/{client-Bc8LANCG.d.ts → client-DCc7a8ZC.d.ts} +2 -2
  23. package/dist/{client-Bc8LANCG.d.ts.map → client-DCc7a8ZC.d.ts.map} +1 -1
  24. package/dist/contract/index.d.ts +8 -8
  25. package/dist/contract/index.js +10 -279
  26. package/dist/contract/index.js.map +1 -1
  27. package/dist/{default-registry-DYB4aHVn.d.ts → default-registry-DMxeb4GD.d.ts} +3 -3
  28. package/dist/{default-registry-DYB4aHVn.d.ts.map → default-registry-DMxeb4GD.d.ts.map} +1 -1
  29. package/dist/{default-registry-CpLko4jB.js → default-registry-xiHrldW8.js} +3 -2
  30. package/dist/{default-registry-CpLko4jB.js.map → default-registry-xiHrldW8.js.map} +1 -1
  31. package/dist/define-agent-eval-C0EqJ9Gz.js +813 -0
  32. package/dist/define-agent-eval-C0EqJ9Gz.js.map +1 -0
  33. package/dist/{define-agent-eval-DE67ECH1.d.ts → define-agent-eval-CGBwPPaq.d.ts} +5 -5
  34. package/dist/{define-agent-eval-DE67ECH1.d.ts.map → define-agent-eval-CGBwPPaq.d.ts.map} +1 -1
  35. package/dist/diagnosis/index.d.ts +319 -0
  36. package/dist/diagnosis/index.d.ts.map +1 -0
  37. package/dist/diagnosis/index.js +1328 -0
  38. package/dist/diagnosis/index.js.map +1 -0
  39. package/dist/{dspy-rlm-engine-Da-l3CpD.js → dspy-rlm-engine-BE9lCzzL.js} +2 -2
  40. package/dist/{dspy-rlm-engine-Da-l3CpD.js.map → dspy-rlm-engine-BE9lCzzL.js.map} +1 -1
  41. package/dist/{engine-D1ykrAKR.d.ts → engine-DNlKhIm0.d.ts} +3 -75
  42. package/dist/engine-DNlKhIm0.d.ts.map +1 -0
  43. package/dist/equal-terms-Blx9Zk5I.js +25 -0
  44. package/dist/equal-terms-Blx9Zk5I.js.map +1 -0
  45. package/dist/experiment/index.d.ts +2 -2
  46. package/dist/experiment/index.js +4 -3
  47. package/dist/experiment/index.js.map +1 -1
  48. package/dist/hosted/index.d.ts +2 -2
  49. package/dist/{index-ww8736Am.d.ts → index-BRsrfSHT.d.ts} +4 -4
  50. package/dist/{index-ww8736Am.d.ts.map → index-BRsrfSHT.d.ts.map} +1 -1
  51. package/dist/{index-DVrtLO5j.d.ts → index-C8jCAOTT.d.ts} +3 -3
  52. package/dist/{index-DVrtLO5j.d.ts.map → index-C8jCAOTT.d.ts.map} +1 -1
  53. package/dist/{index-58wex7KV.d.ts → index-YZrbnBSV.d.ts} +8 -8
  54. package/dist/{index-58wex7KV.d.ts.map → index-YZrbnBSV.d.ts.map} +1 -1
  55. package/dist/index.d.ts +171 -13
  56. package/dist/index.d.ts.map +1 -1
  57. package/dist/index.js +297 -12
  58. package/dist/index.js.map +1 -1
  59. package/dist/{insight-report-C3Bfl762.d.ts → insight-report-DpOo-3iL.d.ts} +2 -2
  60. package/dist/{insight-report-C3Bfl762.d.ts.map → insight-report-DpOo-3iL.d.ts.map} +1 -1
  61. package/dist/{kind-factory-DE088Zhv.js → kind-factory-V79Hw0AU.js} +4 -82
  62. package/dist/kind-factory-V79Hw0AU.js.map +1 -0
  63. package/dist/{llm-judge-C5wbADA5.js → llm-judge-BVso5x-u.js} +5 -3
  64. package/dist/{llm-judge-C5wbADA5.js.map → llm-judge-BVso5x-u.js.map} +1 -1
  65. package/dist/openapi.json +1 -1
  66. package/dist/otel-spans-CQY1pNiN.js +275 -0
  67. package/dist/otel-spans-CQY1pNiN.js.map +1 -0
  68. package/dist/otlp-attributes-BpC_blLF.js +83 -0
  69. package/dist/otlp-attributes-BpC_blLF.js.map +1 -0
  70. package/dist/{run-record-BhMVsHR1.js → paired-promotion-decision-6cn3bLc3.js} +10 -217
  71. package/dist/paired-promotion-decision-6cn3bLc3.js.map +1 -0
  72. package/dist/{paired-promotion-decision-DPsMQm-0.d.ts → paired-promotion-decision-DPKz8ZnF.d.ts} +13 -1
  73. package/dist/{paired-promotion-decision-DPsMQm-0.d.ts.map → paired-promotion-decision-DPKz8ZnF.d.ts.map} +1 -1
  74. package/dist/prime-bridge-transport-Z3cpKGoV.js +43 -0
  75. package/dist/prime-bridge-transport-Z3cpKGoV.js.map +1 -0
  76. package/dist/prime-protocol-CDBcL9NM.d.ts +331 -0
  77. package/dist/prime-protocol-CDBcL9NM.d.ts.map +1 -0
  78. package/dist/{prime-protocol-6tZTVsWm.js → prime-protocol-DcAK9pC8.js} +2 -23
  79. package/dist/prime-protocol-DcAK9pC8.js.map +1 -0
  80. package/dist/{produced-state-D3uWsY8t.js → produced-state-Cn9ECr4v.js} +3 -3
  81. package/dist/{produced-state-D3uWsY8t.js.map → produced-state-Cn9ECr4v.js.map} +1 -1
  82. package/dist/{promotion-policy-CD-bUeyR.js → promotion-policy-CfMNUro2.js} +3 -3
  83. package/dist/{promotion-policy-CD-bUeyR.js.map → promotion-policy-CfMNUro2.js.map} +1 -1
  84. package/dist/{release-confidence-DsETiO4f.d.ts → release-confidence-CSp3MsU4.d.ts} +2 -2
  85. package/dist/{release-confidence-DsETiO4f.d.ts.map → release-confidence-CSp3MsU4.d.ts.map} +1 -1
  86. package/dist/reporting.d.ts +3 -3
  87. package/dist/{researcher-CJkH-6Ty.d.ts → researcher-D7E92STT.d.ts} +2 -2
  88. package/dist/{researcher-CJkH-6Ty.d.ts.map → researcher-D7E92STT.d.ts.map} +1 -1
  89. package/dist/rl.d.ts +2 -2
  90. package/dist/rl.js +2 -1
  91. package/dist/rl.js.map +1 -1
  92. package/dist/run-record-CNNzWGpB.js +216 -0
  93. package/dist/run-record-CNNzWGpB.js.map +1 -0
  94. package/dist/{semantic-concept-judge-5r9hDEBA.js → semantic-concept-judge-Cnt0TWp0.js} +3 -3
  95. package/dist/{semantic-concept-judge-5r9hDEBA.js.map → semantic-concept-judge-Cnt0TWp0.js.map} +1 -1
  96. package/dist/{sequential-BfDkCBGo.js → sequential-lwJOTRON.js} +2 -2
  97. package/dist/{sequential-BfDkCBGo.js.map → sequential-lwJOTRON.js.map} +1 -1
  98. package/dist/{skillopt-optimization-method-B2zxoDk0.js → skillopt-optimization-method-Cc9SVLqi.js} +4 -3
  99. package/dist/{skillopt-optimization-method-B2zxoDk0.js.map → skillopt-optimization-method-Cc9SVLqi.js.map} +1 -1
  100. package/dist/{statistical-heldout-Bu4oNOJh.d.ts → statistical-heldout-DTmhTW8P.d.ts} +2 -2
  101. package/dist/{statistical-heldout-Bu4oNOJh.d.ts.map → statistical-heldout-DTmhTW8P.d.ts.map} +1 -1
  102. package/dist/statistical-heldout-GcLIPxk4.js +256 -0
  103. package/dist/statistical-heldout-GcLIPxk4.js.map +1 -0
  104. package/dist/{store-otlp-rU5PFlOB.js → store-otlp-2gYCy-v0.js} +2 -2
  105. package/dist/{store-otlp-rU5PFlOB.js.map → store-otlp-2gYCy-v0.js.map} +1 -1
  106. package/dist/{store-tool-spans-zrtftcRe.js → store-tool-spans-B1e6Auc2.js} +4 -3
  107. package/dist/{store-tool-spans-zrtftcRe.js.map → store-tool-spans-B1e6Auc2.js.map} +1 -1
  108. package/dist/{store-tool-spans-DwdokbHg.d.ts → store-tool-spans-B3bX9yqn.d.ts} +3 -3
  109. package/dist/{store-tool-spans-DwdokbHg.d.ts.map → store-tool-spans-B3bX9yqn.d.ts.map} +1 -1
  110. package/dist/{summary-report-BzFS5UhZ.d.ts → summary-report-Iy6kXo8x.d.ts} +2 -2
  111. package/dist/{summary-report-BzFS5UhZ.d.ts.map → summary-report-Iy6kXo8x.d.ts.map} +1 -1
  112. package/dist/{task-failure-attributes-BOxjfbhr.js → task-failure-attributes-B34TiMBe.js} +2 -2
  113. package/dist/{task-failure-attributes-BOxjfbhr.js.map → task-failure-attributes-B34TiMBe.js.map} +1 -1
  114. package/dist/{tool-groups-qHMydbAH.d.ts → tool-groups-veQYPiTy.d.ts} +2 -2
  115. package/dist/tool-groups-veQYPiTy.d.ts.map +1 -0
  116. package/dist/trace-repair/index.d.ts +1 -1
  117. package/dist/trace-repair/index.js +2 -1
  118. package/dist/trace-repair/index.js.map +1 -1
  119. package/dist/traces.d.ts +2 -2
  120. package/dist/traces.js +5 -4
  121. package/dist/traces.js.map +1 -1
  122. package/docs/design/statistics-decisions.md +6 -0
  123. package/docs/diagnosis.md +42 -0
  124. package/docs/public-api.md +2 -0
  125. package/package.json +7 -1
  126. package/dist/benchmark-command-DlSePljH.js.map +0 -1
  127. package/dist/campaign-evidence-aWEmfROB.js.map +0 -1
  128. package/dist/define-agent-eval-RTekxhUE.js.map +0 -1
  129. package/dist/engine-D1ykrAKR.d.ts.map +0 -1
  130. package/dist/kind-factory-DE088Zhv.js.map +0 -1
  131. package/dist/prime-protocol-6tZTVsWm.js.map +0 -1
  132. package/dist/run-record-BhMVsHR1.js.map +0 -1
  133. package/dist/tool-groups-qHMydbAH.d.ts.map +0 -1
@@ -1,8 +1,8 @@
1
1
  import { D as TraceAnalysisStore, i as AnalystFinding, n as AnalystContext } from "./types-BjB2wZM9.js";
2
2
  import { l as ExactCapableAnalyst } from "./exact-types-BEHN7tnq.js";
3
3
  import { n as AnalystRegistry, r as AnalystRegistryOptions } from "./registry-D5Ly3GRV.js";
4
- import { a as TraceAnalystLimits, i as TraceAnalysisEngineResult, l as RawAnalystFinding, n as TraceAnalysisEngine } from "./engine-D1ykrAKR.js";
5
- import { t as TraceToolGroupName } from "./tool-groups-qHMydbAH.js";
4
+ import { a as TraceAnalystLimits, i as TraceAnalysisEngineResult, l as RawAnalystFinding, n as TraceAnalysisEngine } from "./engine-DNlKhIm0.js";
5
+ import { t as TraceToolGroupName } from "./tool-groups-veQYPiTy.js";
6
6
  //#region src/trace-analyst/behavioral-metrics.d.ts
7
7
  type SuboptimalCode = 'monotonic-input-growth' | 'output-length-decay' | 'single-tool-dependency' | 'no-self-verification';
8
8
  interface SuboptimalSignal {
@@ -110,4 +110,4 @@ interface DefaultAnalystRegistryOptions {
110
110
  declare function buildDefaultAnalystRegistry(options?: DefaultAnalystRegistryOptions): AnalystRegistry;
111
111
  //#endregion
112
112
  export { createTraceAnalyst as a, BehavioralAnalystOptions as c, TraceAnalystDefinition as i, behavioralAnalyst as l, buildDefaultAnalystRegistry as n, renderPriorFindings as o, CreateTraceAnalystOptions as r, runTraceAnalyst as s, DefaultAnalystRegistryOptions as t, deriveEfficiencyFindings as u };
113
- //# sourceMappingURL=default-registry-DYB4aHVn.d.ts.map
113
+ //# sourceMappingURL=default-registry-DMxeb4GD.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"default-registry-DYB4aHVn.d.ts","names":[],"sources":["../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/kind-factory.ts","../src/analyst/default-registry.ts"],"mappings":";;;;;;KAoBY;UAMK;EACf,MAAM;EACN;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;EACA;;EAEA,gBAAgB;;EAEhB;EACA;EACA,eAAe;EACf;EACA;;EAEA;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,sBAAsB;EACtB,uBAAuB;;;;UC1BR;;EAEf;;EAEA;;;;;;iBA+Dc,yBACd,SAAS,mBACT;EAAQ;EAAoB;IAC3B;;iBAkCa,kBACd,UAAS,2BACR,oBAAoB;;;;UCrHN;EACf;EACA;EACA;EACA;;EAEA,sBAAsB,SAAS;;EAE/B;;EAEA,WAAW;;EAEX,kBACE,OAAO,oBACP,SAAS,wCACe;EAC1B,SAAS,QAAQ;EACjB,eAAe,KAAK,mBAAmB,SAAS,mBAAmB;EACnE;EACA;;UAGe;EACf,QAAQ;;EAER;EACA;;;iBAIoB,gBAAgB;EACpC,YAAY;EACZ,QAAQ;EACR,OAAO;EACP,SAAS;EACT;IACE,QAAQ;;iBAgHI,mBACd,YAAY,wBACZ,SAAS,4BACR,oBAAoB;iBA2RP,oBAAoB,OAAO;;;UChc1B;;;;;;EAMf,SAAS;EACT,uBAAuB;;EAEvB;EACA,aAAa;EACb,WAAW;;iBAGG,4BACd,UAAS,gCACR"}
1
+ {"version":3,"file":"default-registry-DMxeb4GD.d.ts","names":[],"sources":["../src/trace-analyst/behavioral-metrics.ts","../src/analyst/behavioral-analyst.ts","../src/analyst/kind-factory.ts","../src/analyst/default-registry.ts"],"mappings":";;;;;;KAoBY;UAMK;EACf,MAAM;EACN;;EAEA;;EAEA,UAAU;;UAGK;;EAEf;EACA;;EAEA,gBAAgB;;EAEhB;EACA;EACA,eAAe;EACf;EACA;;EAEA;EACA;EACA,SAAS;;UAGM;EACf;EACA;EACA,sBAAsB;EACtB,uBAAuB;;;;UC1BR;;EAEf;;EAEA;;;;;;iBA+Dc,yBACd,SAAS,mBACT;EAAQ;EAAoB;IAC3B;;iBAkCa,kBACd,UAAS,2BACR,oBAAoB;;;;UCrHN;EACf;EACA;EACA;EACA;;EAEA,sBAAsB,SAAS;;EAE/B;;EAEA,WAAW;;EAEX,kBACE,OAAO,oBACP,SAAS,wCACe;EAC1B,SAAS,QAAQ;EACjB,eAAe,KAAK,mBAAmB,SAAS,mBAAmB;EACnE;EACA;;UAGe;EACf,QAAQ;;EAER;EACA;;;iBAIoB,gBAAgB;EACpC,YAAY;EACZ,QAAQ;EACR,OAAO;EACP,SAAS;EACT;IACE,QAAQ;;iBAgHI,mBACd,YAAY,wBACZ,SAAS,4BACR,oBAAoB;iBA2RP,oBAAoB,OAAO;;;UChc1B;;;;;;EAMf,SAAS;EACT,uBAAuB;;EAEvB;EACA,aAAa;EACb,WAAW;;iBAGG,4BACd,UAAS,gCACR"}
@@ -1,9 +1,10 @@
1
1
  import { a as hashCanonical, r as canonicalString } from "./canonical-BsZVgBhW.js";
2
2
  import { t as combineAbortSignals } from "./abort-signal-CtzAM_sJ.js";
3
3
  import { a as deepFreezeCanonicalJson, r as makeFinding } from "./types-C8VBbiw6.js";
4
- import { B as findingSubjectGrammarPromptFor, X as snapshotExactExecutionPlan, Y as snapshotExactExecutionComponentIdentity, k as spanEpochMillis, t as createTraceAnalyst } from "./kind-factory-DE088Zhv.js";
4
+ import { B as findingSubjectGrammarPromptFor, G as snapshotExactExecutionComponentIdentity, K as snapshotExactExecutionPlan, k as spanEpochMillis, t as createTraceAnalyst } from "./kind-factory-V79Hw0AU.js";
5
5
  import { i as validateUsageSettlementTimeout, t as assertValidAnalystUsageReceipt } from "./usage-receipt-Cx3xNeyB.js";
6
6
  import { LLM_CONTEXT_TOKENS, LLM_INPUT_TOKEN_ATTR_KEYS, LLM_OUTPUT_TOKEN_ATTR_KEYS, TOOL_NAME_ATTR_KEYS } from "./trace-attributes.js";
7
+ import "./otlp-attributes-BpC_blLF.js";
7
8
  import { t as executionTrackByLane } from "./execution-tracks-CpgFPpS5.js";
8
9
  import { t as analyzeSupervisorRunIntegrity } from "./integrity-CgQce0FV.js";
9
10
  import { randomUUID } from "node:crypto";
@@ -2364,4 +2365,4 @@ function buildDefaultAnalystRegistry(options = {}) {
2364
2365
  //#endregion
2365
2366
  export { completedAnalystReviewQuality as _, KNOWLEDGE_POISONING_KIND_SPEC as a, validateAnalystReviewDecisions as b, FAILURE_MODE_KIND_SPEC as c, emitControlIntegrityFindings as d, behavioralAnalyst as f, assertUniqueFindingIds as g, analystRunDigest as h, DEFAULT_TRACE_ANALYST_KINDS as i, CONTROL_INTEGRITY_ANALYST as l, analystFindingDigest as m, AnalystRegistry as n, KNOWLEDGE_GAP_KIND_SPEC as o, deriveEfficiencyFindings as p, ExactAnalystRunExecutionError as r, IMPROVEMENT_KIND_SPEC as s, buildDefaultAnalystRegistry as t, ControlIntegrityAnalyst as u, readAnalystReview as v, snapshotAnalystRun as y };
2366
2367
 
2367
- //# sourceMappingURL=default-registry-CpLko4jB.js.map
2368
+ //# sourceMappingURL=default-registry-xiHrldW8.js.map