@tangle-network/agent-eval 0.109.1 → 0.110.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
Files changed (177) hide show
  1. package/CHANGELOG.md +9 -0
  2. package/dist/analyst/index.d.ts +10 -12
  3. package/dist/analyst/index.js +8 -11
  4. package/dist/analyst/index.js.map +1 -1
  5. package/dist/{analyze-runs-DJYpep3L.d.ts → analyze-runs-Dmz6LA9e.d.ts} +4 -4
  6. package/dist/{baseline-Bbid3WoO.d.ts → baseline-DsNteOgR.d.ts} +32 -2
  7. package/dist/belief-state/index.d.ts +6 -6
  8. package/dist/benchmarks/index.d.ts +4 -4
  9. package/dist/benchmarks/index.js +7 -8
  10. package/dist/builder-eval/index.d.ts +4 -4
  11. package/dist/builder-eval/index.js +1 -2
  12. package/dist/builder-eval/index.js.map +1 -1
  13. package/dist/{calibration-BPmzuVPk.d.ts → calibration-Dz8TQV4y.d.ts} +2 -2
  14. package/dist/campaign/index.d.ts +62 -20
  15. package/dist/campaign/index.js +9 -8
  16. package/dist/{chunk-LOJ2QVCE.js → chunk-2IY4ILP4.js} +2 -2
  17. package/dist/{chunk-LIEJUH2I.js → chunk-6PL5MGDL.js} +9 -9
  18. package/dist/{chunk-2OGPXHOB.js → chunk-7NX6ZSBG.js} +36 -7
  19. package/dist/chunk-7NX6ZSBG.js.map +1 -0
  20. package/dist/{chunk-R6D7NEYJ.js → chunk-GBI5J5DB.js} +81 -11
  21. package/dist/chunk-GBI5J5DB.js.map +1 -0
  22. package/dist/{chunk-YEHAEDUD.js → chunk-IMWDSFUM.js} +604 -2
  23. package/dist/chunk-IMWDSFUM.js.map +1 -0
  24. package/dist/{chunk-OVPVM4JC.js → chunk-J4AKLZEV.js} +15 -4
  25. package/dist/{chunk-OVPVM4JC.js.map → chunk-J4AKLZEV.js.map} +1 -1
  26. package/dist/{chunk-JZXGWLK5.js → chunk-MHNQWM4I.js} +62 -6
  27. package/dist/chunk-MHNQWM4I.js.map +1 -0
  28. package/dist/{chunk-QRVS7MX4.js → chunk-OW47B5WA.js} +3 -5
  29. package/dist/{chunk-QRVS7MX4.js.map → chunk-OW47B5WA.js.map} +1 -1
  30. package/dist/{chunk-DBDRR6GF.js → chunk-PLOMR3HP.js} +48 -2
  31. package/dist/chunk-PLOMR3HP.js.map +1 -0
  32. package/dist/{chunk-GDZAWO2I.js → chunk-QFGTU7MT.js} +2 -2
  33. package/dist/{chunk-V7HNA47Z.js → chunk-RSVSSZKF.js} +5 -5
  34. package/dist/{chunk-5PK3626Q.js → chunk-XRGOKCMO.js} +88 -17
  35. package/dist/chunk-XRGOKCMO.js.map +1 -0
  36. package/dist/{code-agent-session-rnJKlqmT.d.ts → code-agent-session-yitf9I-F.d.ts} +1 -1
  37. package/dist/contract/index.d.ts +26 -28
  38. package/dist/contract/index.js +11 -13
  39. package/dist/contract/index.js.map +1 -1
  40. package/dist/{control-B8UthSBL.d.ts → control-U8LBKUES.d.ts} +5 -6
  41. package/dist/control.d.ts +8 -9
  42. package/dist/control.js +6 -8
  43. package/dist/{dataset-DS7ytHZU.d.ts → dataset-NENEzRgk.d.ts} +1 -1
  44. package/dist/{default-registry-BswHCXnU.d.ts → default-registry-Bcf1uKVI.d.ts} +1 -2
  45. package/dist/{emitter-C2rqGH_l.d.ts → emitter-BRchAAAx.d.ts} +2 -2
  46. package/dist/{failure-cluster-DH9Flgcf.d.ts → failure-cluster-C48PiReX.d.ts} +2 -2
  47. package/dist/feedback-trajectory-pDcz1lQ1.d.ts +348 -0
  48. package/dist/{gepa-B3x5Ulcv.d.ts → gepa-T8T215nw.d.ts} +149 -6
  49. package/dist/hosted/index.d.ts +7 -7
  50. package/dist/{index-pPtfoIJO.d.ts → index-Dc3VLGhp.d.ts} +2 -2
  51. package/dist/index.d.ts +645 -61
  52. package/dist/index.js +1282 -190
  53. package/dist/index.js.map +1 -1
  54. package/dist/{insight-report-B4xrdwEK.d.ts → insight-report-D4cXFsLt.d.ts} +1 -1
  55. package/dist/{integrity-DqGZg3st.d.ts → integrity-qemeBAyx.d.ts} +1 -1
  56. package/dist/{types-D1ytG0Yg.d.ts → kind-factory-20hcaYpf.d.ts} +169 -2
  57. package/dist/meta-eval/index.d.ts +5 -5
  58. package/dist/meta-eval/index.js +1 -2
  59. package/dist/meta-eval/index.js.map +1 -1
  60. package/dist/{multi-layer-verifier-CI4jdX-q.d.ts → multi-layer-verifier-BsqKuLyN.d.ts} +1 -1
  61. package/dist/multishot/index.d.ts +3 -3
  62. package/dist/openapi.json +1 -1
  63. package/dist/pipelines/index.d.ts +6 -7
  64. package/dist/pipelines/index.js +3 -6
  65. package/dist/pipelines/index.js.map +1 -1
  66. package/dist/{policy-edit-DQUXYMDm.d.ts → policy-edit-D2bBDZDf.d.ts} +2 -2
  67. package/dist/{pre-registration-BUhVPzE7.d.ts → pre-registration-BepVVa6P.d.ts} +3 -3
  68. package/dist/{provenance-DdDhf6cg.d.ts → provenance-CyxkvEi9.d.ts} +3 -5
  69. package/dist/{query-0aTmbmQe.d.ts → query-Ck190MOd.d.ts} +2 -2
  70. package/dist/{release-report-DeJpsBiA.d.ts → release-report-oBfOz8ku.d.ts} +3 -3
  71. package/dist/reporting.d.ts +8 -8
  72. package/dist/{researcher-Wc7dx6GM.d.ts → researcher-CaH0CwFC.d.ts} +6 -6
  73. package/dist/rl.d.ts +568 -15
  74. package/dist/rl.js +4 -4
  75. package/dist/{rubric-predictive-validity-DPnyG-CE.d.ts → rubric-predictive-validity-C-fMteAW.d.ts} +1 -1
  76. package/dist/{run-record-I-Z3JNvO.d.ts → run-record-DksGsfgv.d.ts} +1 -1
  77. package/dist/{runtime-trajectory-iW9IhV3e.d.ts → runtime-trajectory-h5i0SZUj.d.ts} +1 -1
  78. package/dist/{schema-m0gsnbt3.d.ts → schema-SGWcK9wa.d.ts} +1 -1
  79. package/dist/{semantic-concept-judge-BmNZPB_j.d.ts → semantic-concept-judge-D7z6JCLZ.d.ts} +57 -4
  80. package/dist/{store-BcFXE6LG.d.ts → store-BsVi7ncX.d.ts} +1 -1
  81. package/dist/storyboard/index.d.ts +1 -1
  82. package/dist/{summary-report-QMZVe3P-.d.ts → summary-report-Bz-0-t8v.d.ts} +2 -2
  83. package/dist/{test-graded-scenario-DeODGLra.d.ts → test-graded-scenario-mzYBKspu.d.ts} +3 -3
  84. package/dist/traces.d.ts +54 -11
  85. package/dist/traces.js +25 -27
  86. package/dist/{types-BdIv5dvA.d.ts → types-v--ctu-b.d.ts} +2 -2
  87. package/dist/wire/index.d.ts +5 -6
  88. package/docs/improvement-glossary.md +14 -13
  89. package/package.json +1 -71
  90. package/dist/adapters/http.d.ts +0 -142
  91. package/dist/adapters/http.js +0 -203
  92. package/dist/adapters/http.js.map +0 -1
  93. package/dist/adapters/langchain.d.ts +0 -95
  94. package/dist/adapters/langchain.js +0 -34
  95. package/dist/adapters/langchain.js.map +0 -1
  96. package/dist/adapters/otel.d.ts +0 -112
  97. package/dist/adapters/otel.js +0 -110
  98. package/dist/adapters/otel.js.map +0 -1
  99. package/dist/chunk-2OGPXHOB.js.map +0 -1
  100. package/dist/chunk-45EEMHTC.js +0 -35
  101. package/dist/chunk-45EEMHTC.js.map +0 -1
  102. package/dist/chunk-5BKGXME7.js +0 -65
  103. package/dist/chunk-5BKGXME7.js.map +0 -1
  104. package/dist/chunk-5PK3626Q.js.map +0 -1
  105. package/dist/chunk-6SK5VFYK.js +0 -100
  106. package/dist/chunk-6SK5VFYK.js.map +0 -1
  107. package/dist/chunk-DBDRR6GF.js.map +0 -1
  108. package/dist/chunk-DJWX3GVS.js +0 -81
  109. package/dist/chunk-DJWX3GVS.js.map +0 -1
  110. package/dist/chunk-FOUG2VVS.js +0 -855
  111. package/dist/chunk-FOUG2VVS.js.map +0 -1
  112. package/dist/chunk-JZXGWLK5.js.map +0 -1
  113. package/dist/chunk-K7QEIHHJ.js +0 -613
  114. package/dist/chunk-K7QEIHHJ.js.map +0 -1
  115. package/dist/chunk-KKHDIONI.js +0 -414
  116. package/dist/chunk-KKHDIONI.js.map +0 -1
  117. package/dist/chunk-KMPRBJK4.js +0 -74
  118. package/dist/chunk-KMPRBJK4.js.map +0 -1
  119. package/dist/chunk-Q2JRAWRI.js +0 -196
  120. package/dist/chunk-Q2JRAWRI.js.map +0 -1
  121. package/dist/chunk-R6D7NEYJ.js.map +0 -1
  122. package/dist/chunk-RZTMDUO7.js +0 -49
  123. package/dist/chunk-RZTMDUO7.js.map +0 -1
  124. package/dist/chunk-STGVSCDH.js +0 -202
  125. package/dist/chunk-STGVSCDH.js.map +0 -1
  126. package/dist/chunk-YEHAEDUD.js.map +0 -1
  127. package/dist/control-runtime-Acf9CGhw.d.ts +0 -182
  128. package/dist/corpus-eBVwhCp1.d.ts +0 -560
  129. package/dist/counterfactual-DlOz8PBx.d.ts +0 -85
  130. package/dist/diagnose.d.ts +0 -252
  131. package/dist/diagnose.js +0 -382
  132. package/dist/diagnose.js.map +0 -1
  133. package/dist/feedback-trajectory-C9KCo8ag.d.ts +0 -169
  134. package/dist/governance/index.d.ts +0 -135
  135. package/dist/governance/index.js +0 -18
  136. package/dist/governance/index.js.map +0 -1
  137. package/dist/groundedness/index.d.ts +0 -112
  138. package/dist/groundedness/index.js +0 -77
  139. package/dist/groundedness/index.js.map +0 -1
  140. package/dist/harness-optimizer-mOl9XX_O.d.ts +0 -106
  141. package/dist/kind-factory-DvIGo_cP.d.ts +0 -171
  142. package/dist/knowledge/index.d.ts +0 -103
  143. package/dist/knowledge/index.js +0 -18
  144. package/dist/knowledge/index.js.map +0 -1
  145. package/dist/pareto-E-pembql.d.ts +0 -81
  146. package/dist/perf/index.d.ts +0 -123
  147. package/dist/perf/index.js +0 -18
  148. package/dist/perf/index.js.map +0 -1
  149. package/dist/prm/index.d.ts +0 -104
  150. package/dist/prm/index.js +0 -265
  151. package/dist/prm/index.js.map +0 -1
  152. package/dist/product-benchmark/index.d.ts +0 -247
  153. package/dist/product-benchmark/index.js +0 -37
  154. package/dist/product-benchmark/index.js.map +0 -1
  155. package/dist/red-team-KmmiqBlY.d.ts +0 -63
  156. package/dist/redact-B40YG2M_.d.ts +0 -45
  157. package/dist/rubric-Cc6UHvUb.d.ts +0 -73
  158. package/dist/run-critic-CmMf05uV.d.ts +0 -56
  159. package/dist/sink-fetch-B1Yg4Til.d.ts +0 -101
  160. package/dist/telemetry/file.d.ts +0 -19
  161. package/dist/telemetry/file.js +0 -45
  162. package/dist/telemetry/file.js.map +0 -1
  163. package/dist/telemetry/index.d.ts +0 -38
  164. package/dist/telemetry/index.js +0 -130
  165. package/dist/telemetry/index.js.map +0 -1
  166. package/dist/testing-C21CHsq2.d.ts +0 -20
  167. package/dist/testing.d.ts +0 -1
  168. package/dist/testing.js +0 -8
  169. package/dist/testing.js.map +0 -1
  170. package/dist/trajectory-2TkpSEVh.d.ts +0 -33
  171. package/dist/workflow/index.d.ts +0 -496
  172. package/dist/workflow/index.js +0 -2178
  173. package/dist/workflow/index.js.map +0 -1
  174. /package/dist/{chunk-LOJ2QVCE.js.map → chunk-2IY4ILP4.js.map} +0 -0
  175. /package/dist/{chunk-LIEJUH2I.js.map → chunk-6PL5MGDL.js.map} +0 -0
  176. /package/dist/{chunk-GDZAWO2I.js.map → chunk-QFGTU7MT.js.map} +0 -0
  177. /package/dist/{chunk-V7HNA47Z.js.map → chunk-RSVSSZKF.js.map} +0 -0
@@ -10,7 +10,11 @@ import {
10
10
  renderAnalystEvidence,
11
11
  runImprovementLoop,
12
12
  surfaceHash
13
- } from "./chunk-OVPVM4JC.js";
13
+ } from "./chunk-J4AKLZEV.js";
14
+ import {
15
+ estimateCost,
16
+ isModelPriced
17
+ } from "./chunk-VI2UW6B6.js";
14
18
  import {
15
19
  assertRealBackend,
16
20
  contentHash,
@@ -18,25 +22,18 @@ import {
18
22
  runCampaign,
19
23
  summarizeBackendIntegrity
20
24
  } from "./chunk-3PFZBGMR.js";
21
- import {
22
- estimateCost,
23
- isModelPriced
24
- } from "./chunk-VI2UW6B6.js";
25
25
  import {
26
26
  admitPolicyEdit,
27
27
  applyPolicyEditToSurface,
28
28
  clamp01,
29
29
  isPolicyEdit,
30
30
  policyEditsFromFindings
31
- } from "./chunk-LOJ2QVCE.js";
31
+ } from "./chunk-2IY4ILP4.js";
32
32
  import {
33
33
  AnalystRegistry,
34
34
  DEFAULT_TRACE_ANALYST_KINDS,
35
35
  createTraceAnalystKind
36
- } from "./chunk-2OGPXHOB.js";
37
- import {
38
- callLlm
39
- } from "./chunk-FUCQVFMU.js";
36
+ } from "./chunk-7NX6ZSBG.js";
40
37
  import {
41
38
  eProcess,
42
39
  mulberry32,
@@ -59,6 +56,9 @@ import {
59
56
  import {
60
57
  canonicalize
61
58
  } from "./chunk-VSMTAMNK.js";
59
+ import {
60
+ callLlm
61
+ } from "./chunk-FUCQVFMU.js";
62
62
  import {
63
63
  AgentEvalError,
64
64
  JudgeError,
@@ -3870,6 +3870,75 @@ ${block}` : block;
3870
3870
  };
3871
3871
  }
3872
3872
 
3873
+ // src/campaign/proposers/composite.ts
3874
+ function compositeProposer(opts) {
3875
+ const members = opts.proposers;
3876
+ if (members.length === 0)
3877
+ throw new Error("compositeProposer: at least one member proposer required");
3878
+ const weights = opts.weights ?? members.map(() => 1);
3879
+ if (weights.length !== members.length || weights.some((w) => !(w > 0))) {
3880
+ throw new Error("compositeProposer: weights must match proposers length and be positive");
3881
+ }
3882
+ return {
3883
+ kind: `composite(${members.map((m) => m.kind).join("+")})`,
3884
+ async propose(ctx) {
3885
+ const totalWeight = weights.reduce((a, b) => a + b, 0);
3886
+ const exact = weights.map((w) => ctx.populationSize * w / totalWeight);
3887
+ const shares = exact.map((e) => Math.floor(e));
3888
+ let remaining = ctx.populationSize - shares.reduce((a, b) => a + b, 0);
3889
+ const byRemainder = exact.map((e, i) => ({ i, rem: e - Math.floor(e) })).sort((a, b) => b.rem - a.rem);
3890
+ for (const { i } of byRemainder) {
3891
+ if (remaining <= 0) break;
3892
+ shares[i] = (shares[i] ?? 0) + 1;
3893
+ remaining -= 1;
3894
+ }
3895
+ const pool = [];
3896
+ const seen = /* @__PURE__ */ new Set();
3897
+ const errors = [];
3898
+ for (let i = 0; i < members.length; i += 1) {
3899
+ const member = members[i];
3900
+ const share = shares[i] ?? 0;
3901
+ if (!member || share === 0 || ctx.signal.aborted) continue;
3902
+ try {
3903
+ const proposals = await member.propose({ ...ctx, populationSize: share });
3904
+ for (const proposal of proposals) {
3905
+ const isCandidate = typeof proposal === "object" && proposal !== null && "surface" in proposal;
3906
+ const surface = isCandidate ? proposal.surface : proposal;
3907
+ const label = isCandidate ? proposal.label : "candidate";
3908
+ const rationale = isCandidate ? proposal.rationale : "";
3909
+ const key = typeof surface === "string" ? surface : JSON.stringify(surface);
3910
+ if (seen.has(key)) continue;
3911
+ seen.add(key);
3912
+ pool.push({ surface, label: `${member.kind}:${label}`, rationale });
3913
+ }
3914
+ } catch (err) {
3915
+ errors.push(`${member.kind}: ${err instanceof Error ? err.message : String(err)}`);
3916
+ }
3917
+ }
3918
+ if (pool.length === 0) {
3919
+ throw new Error(
3920
+ `compositeProposer: every member failed or proposed nothing${errors.length ? ` \u2014 ${errors.join("; ")}` : ""}`
3921
+ );
3922
+ }
3923
+ if (errors.length > 0) {
3924
+ console.warn(
3925
+ `[compositeProposer] ${errors.length} member(s) failed this generation: ${errors.join("; ")}`
3926
+ );
3927
+ }
3928
+ return pool.slice(0, ctx.populationSize);
3929
+ },
3930
+ decide(args) {
3931
+ const votes = members.filter((m) => typeof m.decide === "function").map((m) => m.decide(args));
3932
+ if (votes.length === 0) return { stop: false };
3933
+ const allStop = votes.every((v) => v.stop);
3934
+ return allStop ? {
3935
+ stop: true,
3936
+ reason: votes.map((v) => v.reason).filter(Boolean).join("; ") || "all members converged"
3937
+ } : { stop: false };
3938
+ }
3939
+ };
3940
+ }
3941
+
3873
3942
  // src/campaign/proposers/halo.ts
3874
3943
  import { execFile } from "child_process";
3875
3944
  import { promisify } from "util";
@@ -4393,6 +4462,7 @@ export {
4393
4462
  ProfileMatrixError,
4394
4463
  runProfileMatrix,
4395
4464
  aceProposer,
4465
+ compositeProposer,
4396
4466
  haloProposer,
4397
4467
  memoryCurationProposer,
4398
4468
  policyEditProposer,
@@ -4405,4 +4475,4 @@ export {
4405
4475
  gitWorktreeAdapter,
4406
4476
  resolveWorktreePath
4407
4477
  };
4408
- //# sourceMappingURL=chunk-R6D7NEYJ.js.map
4478
+ //# sourceMappingURL=chunk-GBI5J5DB.js.map