@tangle-network/agent-eval 0.118.2 → 0.119.0

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 (67) hide show
  1. package/CHANGELOG.md +21 -0
  2. package/README.md +26 -0
  3. package/dist/analyst/index.d.ts +143 -21
  4. package/dist/analyst/index.js +46 -12
  5. package/dist/analyst/index.js.map +1 -1
  6. package/dist/benchmarks/index.d.ts +50 -40
  7. package/dist/benchmarks/index.js +9 -9
  8. package/dist/campaign/index.d.ts +231 -156
  9. package/dist/campaign/index.js +8 -8
  10. package/dist/{chunk-JTMFT5QZ.js → chunk-4I2E3LLO.js} +2 -2
  11. package/dist/{chunk-BUBU3YHT.js → chunk-6QIM2EAP.js} +6 -116
  12. package/dist/chunk-6QIM2EAP.js.map +1 -0
  13. package/dist/{chunk-DZBCDDIE.js → chunk-7A4LIMMY.js} +846 -253
  14. package/dist/chunk-7A4LIMMY.js.map +1 -0
  15. package/dist/{chunk-KK2O4VWA.js → chunk-D7AEXSM5.js} +2 -2
  16. package/dist/{chunk-GCT3DQWU.js → chunk-F6YUH3L4.js} +41 -30
  17. package/dist/{chunk-GCT3DQWU.js.map → chunk-F6YUH3L4.js.map} +1 -1
  18. package/dist/{chunk-63KG3ASG.js → chunk-FIUFRXP3.js} +2 -2
  19. package/dist/{chunk-J23QK2U5.js → chunk-GERDAIAL.js} +8 -8
  20. package/dist/{chunk-J23QK2U5.js.map → chunk-GERDAIAL.js.map} +1 -1
  21. package/dist/{chunk-TVXPJJ7H.js → chunk-H5UD2323.js} +2 -2
  22. package/dist/{chunk-MLHRFWQK.js → chunk-JHOJHHU7.js} +85 -16
  23. package/dist/chunk-JHOJHHU7.js.map +1 -0
  24. package/dist/chunk-K4DBDHLK.js +158 -0
  25. package/dist/chunk-K4DBDHLK.js.map +1 -0
  26. package/dist/{chunk-DRSFTBKU.js → chunk-LMJ2TGWJ.js} +3 -15
  27. package/dist/chunk-LMJ2TGWJ.js.map +1 -0
  28. package/dist/{chunk-EGNRE7VA.js → chunk-PAHNGS65.js} +2 -2
  29. package/dist/{chunk-I42AAGHL.js → chunk-RLCJQ6VZ.js} +5 -5
  30. package/dist/{chunk-7V3QDWNL.js → chunk-XJ7JVCHB.js} +2 -2
  31. package/dist/{chunk-6JQE3YVE.js → chunk-ZYHJNKI3.js} +6 -6
  32. package/dist/cli.js +2 -2
  33. package/dist/contract/index.d.ts +257 -160
  34. package/dist/contract/index.js +11 -11
  35. package/dist/fuzz.d.ts +19 -1
  36. package/dist/fuzz.js +1 -1
  37. package/dist/index.d.ts +170 -64
  38. package/dist/index.js +30 -28
  39. package/dist/index.js.map +1 -1
  40. package/dist/openapi.json +1 -1
  41. package/dist/primeintellect/index.d.ts +311 -0
  42. package/dist/primeintellect/index.js +348 -0
  43. package/dist/primeintellect/index.js.map +1 -0
  44. package/dist/rl.d.ts +4 -0
  45. package/dist/{run-campaign-DWC67KJP.js → run-campaign-OBXSN5WK.js} +3 -3
  46. package/dist/trace-attributes.d.ts +17 -1
  47. package/dist/trace-attributes.js +27 -3
  48. package/dist/traces.d.ts +16 -16
  49. package/dist/traces.js +18 -18
  50. package/dist/wire/index.d.ts +19 -1
  51. package/dist/wire/index.js +2 -2
  52. package/package.json +6 -1
  53. package/dist/chunk-BUBU3YHT.js.map +0 -1
  54. package/dist/chunk-DRSFTBKU.js.map +0 -1
  55. package/dist/chunk-DZBCDDIE.js.map +0 -1
  56. package/dist/chunk-MLHRFWQK.js.map +0 -1
  57. package/dist/chunk-NW4AQBXD.js +0 -53
  58. package/dist/chunk-NW4AQBXD.js.map +0 -1
  59. /package/dist/{chunk-JTMFT5QZ.js.map → chunk-4I2E3LLO.js.map} +0 -0
  60. /package/dist/{chunk-KK2O4VWA.js.map → chunk-D7AEXSM5.js.map} +0 -0
  61. /package/dist/{chunk-63KG3ASG.js.map → chunk-FIUFRXP3.js.map} +0 -0
  62. /package/dist/{chunk-TVXPJJ7H.js.map → chunk-H5UD2323.js.map} +0 -0
  63. /package/dist/{chunk-EGNRE7VA.js.map → chunk-PAHNGS65.js.map} +0 -0
  64. /package/dist/{chunk-I42AAGHL.js.map → chunk-RLCJQ6VZ.js.map} +0 -0
  65. /package/dist/{chunk-7V3QDWNL.js.map → chunk-XJ7JVCHB.js.map} +0 -0
  66. /package/dist/{chunk-6JQE3YVE.js.map → chunk-ZYHJNKI3.js.map} +0 -0
  67. /package/dist/{run-campaign-DWC67KJP.js.map → run-campaign-OBXSN5WK.js.map} +0 -0
package/dist/index.js CHANGED
@@ -9,12 +9,12 @@ import {
9
9
  checkBehavioralCanary,
10
10
  checkCanaries,
11
11
  runBehavioralCanaries
12
- } from "./chunk-7V3QDWNL.js";
12
+ } from "./chunk-XJ7JVCHB.js";
13
13
  import {
14
14
  BENCHMARK_SPLIT_SEED,
15
15
  benchmarks_exports,
16
16
  deterministicSplit
17
- } from "./chunk-KK2O4VWA.js";
17
+ } from "./chunk-D7AEXSM5.js";
18
18
  import {
19
19
  DEFAULT_RULES,
20
20
  buildTrajectory,
@@ -53,7 +53,7 @@ import {
53
53
  pairArms,
54
54
  parseCorrectnessResponse,
55
55
  verifyCompletion
56
- } from "./chunk-GCT3DQWU.js";
56
+ } from "./chunk-F6YUH3L4.js";
57
57
  import {
58
58
  DEFAULT_MUTATION_PRIMITIVES,
59
59
  DEFAULT_RED_TEAM_CORPUS,
@@ -91,7 +91,7 @@ import {
91
91
  scoreRedTeamOutput,
92
92
  surfaceContentHash,
93
93
  toolNamesForRun
94
- } from "./chunk-I42AAGHL.js";
94
+ } from "./chunk-RLCJQ6VZ.js";
95
95
  import {
96
96
  BackendIntegrityError,
97
97
  assertRealBackend,
@@ -101,7 +101,7 @@ import {
101
101
  fileVerdictCache,
102
102
  inMemoryVerdictCache,
103
103
  summarizeBackendIntegrity
104
- } from "./chunk-EGNRE7VA.js";
104
+ } from "./chunk-PAHNGS65.js";
105
105
  import {
106
106
  DEFAULT_COMPLEXITY_WEIGHTS,
107
107
  FindingsStore,
@@ -110,17 +110,16 @@ import {
110
110
  SEMANTIC_CONCEPT_JUDGE_VERSION,
111
111
  SKILL_USAGE_ANALYST,
112
112
  SkillUsageAnalyst,
113
- createAnalystAi,
114
113
  createSemanticConceptJudge,
115
114
  defaultIsMaterial,
116
115
  diffFindings,
117
116
  runSemanticConceptJudge
118
- } from "./chunk-DRSFTBKU.js";
117
+ } from "./chunk-LMJ2TGWJ.js";
119
118
  import {
120
119
  buildDefaultAnalystRegistry,
121
120
  computeTraceMetrics,
122
121
  createChatClient
123
- } from "./chunk-J23QK2U5.js";
122
+ } from "./chunk-GERDAIAL.js";
124
123
  import "./chunk-HHWE3POT.js";
125
124
  import {
126
125
  Mutex
@@ -143,6 +142,7 @@ import {
143
142
  clamp01,
144
143
  computeFindingId,
145
144
  computePolicyEditId,
145
+ createAnalystAi,
146
146
  createTraceAnalystKind,
147
147
  isPolicyEdit,
148
148
  makeFinding,
@@ -151,10 +151,11 @@ import {
151
151
  policyEditFromFinding,
152
152
  policyEditsFromFindings,
153
153
  renderPriorFindings,
154
+ renderUpstreamFindings,
154
155
  scorePolicyEditReadiness,
155
156
  validatePolicyEdit,
156
157
  validatePolicyEditCandidateRecord
157
- } from "./chunk-DZBCDDIE.js";
158
+ } from "./chunk-7A4LIMMY.js";
158
159
  import {
159
160
  allCriticalPassed,
160
161
  controlFailureClassFromVerification,
@@ -264,7 +265,7 @@ import {
264
265
  CostReservationExceededError,
265
266
  costForUsage,
266
267
  modelPriceKey
267
- } from "./chunk-MLHRFWQK.js";
268
+ } from "./chunk-JHOJHHU7.js";
268
269
  import {
269
270
  MODEL_PRICING,
270
271
  MetricsCollector,
@@ -303,8 +304,8 @@ import {
303
304
  scoreTraceInsightReadiness,
304
305
  tokenizeDomainWords,
305
306
  traceAnalystOnRunComplete
306
- } from "./chunk-6JQE3YVE.js";
307
- import "./chunk-TVXPJJ7H.js";
307
+ } from "./chunk-ZYHJNKI3.js";
308
+ import "./chunk-H5UD2323.js";
308
309
  import {
309
310
  extractUsage,
310
311
  extractUsageFromResponse,
@@ -326,7 +327,7 @@ import {
326
327
  TRACE_ANALYST_ACTOR_DESCRIPTION_VERSION,
327
328
  TRACE_ANALYST_SUBAGENT_DESCRIPTION,
328
329
  analyzeTraces
329
- } from "./chunk-63KG3ASG.js";
330
+ } from "./chunk-FIUFRXP3.js";
330
331
  import {
331
332
  DEFAULT_REDACTION_RULES,
332
333
  REDACTION_VERSION,
@@ -335,27 +336,15 @@ import {
335
336
  } from "./chunk-GGE4NNQT.js";
336
337
  import {
337
338
  DEFAULT_TRACE_ANALYST_BUDGETS,
338
- LLM_CACHED_TOKEN_ATTR_KEYS,
339
- LLM_CACHE_WRITE_TOKEN_ATTR_KEYS,
340
- LLM_COST_ATTR_KEYS,
341
- LLM_INPUT_TOKEN_ATTR_KEYS,
342
- LLM_MODEL_ATTR_KEYS,
343
- LLM_OUTPUT_TOKEN_ATTR_KEYS,
344
- LLM_REASONING_TOKEN_ATTR_KEYS,
345
339
  OtlpFileTraceStore,
346
- RUN_COST_ATTR_KEYS,
347
- SPAN_KIND_ATTR_KEYS,
348
340
  SpanNotFoundError,
349
- TOOL_NAME_ATTR_KEYS,
350
341
  TRACE_ANALYST_TRUNCATION_MARKER_PREFIX,
351
342
  TraceFileMissingError,
352
343
  TraceNotFoundError,
353
344
  applyToolSpanOtlpAttributes,
354
- asNumber,
355
345
  asString,
356
346
  buildTraceAnalystTools,
357
347
  extractOtlpAttributes,
358
- firstNumberAttr,
359
348
  firstStringAttr,
360
349
  inferOtlpKind,
361
350
  projectOtlpFlatLine,
@@ -363,7 +352,7 @@ import {
363
352
  stringField,
364
353
  traceAnalystFunctionGroup,
365
354
  traceSpanKindToOpenInferenceKind
366
- } from "./chunk-BUBU3YHT.js";
355
+ } from "./chunk-6QIM2EAP.js";
367
356
  import {
368
357
  RunIntegrityError,
369
358
  assertRunCaptured,
@@ -432,21 +421,33 @@ import {
432
421
  import {
433
422
  INPUT_VALUE,
434
423
  LLM_CACHED_TOKENS,
424
+ LLM_CACHED_TOKEN_ATTR_KEYS,
435
425
  LLM_CACHE_WRITE_TOKENS,
426
+ LLM_CACHE_WRITE_TOKEN_ATTR_KEYS,
436
427
  LLM_CONTEXT_TOKENS,
428
+ LLM_COST_ATTR_KEYS,
437
429
  LLM_COST_USD,
438
430
  LLM_INPUT_TOKENS,
431
+ LLM_INPUT_TOKEN_ATTR_KEYS,
432
+ LLM_MODEL_ATTR_KEYS,
439
433
  LLM_MODEL_NAME,
440
434
  LLM_OUTPUT_TOKENS,
435
+ LLM_OUTPUT_TOKEN_ATTR_KEYS,
441
436
  LLM_REASONING_TOKENS,
437
+ LLM_REASONING_TOKEN_ATTR_KEYS,
442
438
  OPENINFERENCE_SPAN_KIND,
443
439
  OUTPUT_VALUE,
440
+ RUN_COST_ATTR_KEYS,
441
+ SPAN_KIND_ATTR_KEYS,
444
442
  TOOL_ARGS_CAPTURED,
445
443
  TOOL_LATENCY_MS,
446
444
  TOOL_NAME,
445
+ TOOL_NAME_ATTR_KEYS,
447
446
  applyLlmSpanOtlpAttributes,
448
- contextInputTokens
449
- } from "./chunk-NW4AQBXD.js";
447
+ asNumber,
448
+ contextInputTokens,
449
+ firstNumberAttr
450
+ } from "./chunk-K4DBDHLK.js";
450
451
  import {
451
452
  __export
452
453
  } from "./chunk-PZ5AY32C.js";
@@ -12040,6 +12041,7 @@ export {
12040
12041
  renderPriorFindings,
12041
12042
  renderReleaseReport,
12042
12043
  renderSteeringText,
12044
+ renderUpstreamFindings,
12043
12045
  repeatedActionDetector,
12044
12046
  replayFeedbackTrajectories,
12045
12047
  replayFeedbackTrajectory,