@tangle-network/agent-eval 0.130.0 → 0.131.0

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 (40) hide show
  1. package/CHANGELOG.md +19 -0
  2. package/README.md +1 -0
  3. package/dist/analyst/index.js +1 -1
  4. package/dist/benchmarks/index.d.ts +1 -1
  5. package/dist/benchmarks/index.js +1 -1
  6. package/dist/{benchmarks-DviOvUNr.js → benchmarks-SRt_Rwuc.js} +3 -3
  7. package/dist/{benchmarks-DviOvUNr.js.map → benchmarks-SRt_Rwuc.js.map} +1 -1
  8. package/dist/campaign/index.d.ts +2 -2
  9. package/dist/campaign/index.js +2 -2
  10. package/dist/{campaign-CBKZvQ1H.js → campaign-1WfDhAl4.js} +4 -3
  11. package/dist/{campaign-CBKZvQ1H.js.map → campaign-1WfDhAl4.js.map} +1 -1
  12. package/dist/concurrency-MUjT7VjM.js +109 -0
  13. package/dist/concurrency-MUjT7VjM.js.map +1 -0
  14. package/dist/contract/index.d.ts +57 -3
  15. package/dist/contract/index.d.ts.map +1 -1
  16. package/dist/contract/index.js +293 -47
  17. package/dist/contract/index.js.map +1 -1
  18. package/dist/{index-DE5fb3EC.d.ts → index-CD_WZ_Xr.d.ts} +8 -2
  19. package/dist/{index-DE5fb3EC.d.ts.map → index-CD_WZ_Xr.d.ts.map} +1 -1
  20. package/dist/{index-CAPUUKaM.d.ts → index-Em67JBjs.d.ts} +3 -3
  21. package/dist/{index-CAPUUKaM.d.ts.map → index-Em67JBjs.d.ts.map} +1 -1
  22. package/dist/index.d.ts +3 -3
  23. package/dist/index.d.ts.map +1 -1
  24. package/dist/index.js +5 -5
  25. package/dist/ledger-core/index.js +1 -1
  26. package/dist/{ledger-core-DtZz1RG0.js → ledger-core-eqaI3PCD.js} +2 -2
  27. package/dist/{ledger-core-DtZz1RG0.js.map → ledger-core-eqaI3PCD.js.map} +1 -1
  28. package/dist/openapi.json +1 -1
  29. package/dist/{semantic-concept-judge-B6cWNJ2K.js → semantic-concept-judge-b5m3irbR.js} +2 -2
  30. package/dist/{semantic-concept-judge-B6cWNJ2K.js.map → semantic-concept-judge-b5m3irbR.js.map} +1 -1
  31. package/dist/{skillopt-optimization-method-B7wX7XkF.d.ts → skillopt-optimization-method-CWKVTnks.d.ts} +2 -1
  32. package/dist/{skillopt-optimization-method-B7wX7XkF.d.ts.map → skillopt-optimization-method-CWKVTnks.d.ts.map} +1 -1
  33. package/dist/{skillopt-optimization-method-D4ODwFVV.js → skillopt-optimization-method-Dd6b38Ud.js} +9 -3
  34. package/dist/skillopt-optimization-method-Dd6b38Ud.js.map +1 -0
  35. package/docs/campaign-proposers.md +1 -0
  36. package/docs/concepts.md +11 -0
  37. package/package.json +1 -1
  38. package/dist/concurrency-DIxRZF_J.js +0 -85
  39. package/dist/concurrency-DIxRZF_J.js.map +0 -1
  40. package/dist/skillopt-optimization-method-D4ODwFVV.js.map +0 -1
@@ -0,0 +1,109 @@
1
+ //#region src/concurrency.ts
2
+ /**
3
+ * concurrency — small primitives the evolution loop needs.
4
+ *
5
+ * `Mutex` is a zero-dep async lock with FIFO fairness. The evolution loop
6
+ * uses it to serialise checkout/build/commit sequences inside a single
7
+ * pool slot, and to gate concurrent JSONL writers (see
8
+ * `lockedJsonlReferenceReplayStore`).
9
+ *
10
+ * Deliberately minimal — no priority queue, no timeouts. If you need
11
+ * those, swap to `async-mutex` at the call site.
12
+ */
13
+ var Mutex = class {
14
+ locked = false;
15
+ waiters = [];
16
+ async acquire() {
17
+ if (!this.locked) {
18
+ this.locked = true;
19
+ return () => this.release();
20
+ }
21
+ return new Promise((resolve) => {
22
+ this.waiters.push(() => {
23
+ resolve(() => this.release());
24
+ });
25
+ });
26
+ }
27
+ release() {
28
+ const next = this.waiters.shift();
29
+ if (next) next();
30
+ else this.locked = false;
31
+ }
32
+ async runExclusive(fn) {
33
+ const release = await this.acquire();
34
+ try {
35
+ return await fn();
36
+ } finally {
37
+ release();
38
+ }
39
+ }
40
+ /** True iff someone holds the lock right now. Diagnostics only. */
41
+ get isLocked() {
42
+ return this.locked;
43
+ }
44
+ /** Pending waiter count. Diagnostics only. */
45
+ get pending() {
46
+ return this.waiters.length;
47
+ }
48
+ };
49
+ /** Map an integer range with bounded work, cancellation, and first-error cleanup. */
50
+ async function mapConcurrentRange(options) {
51
+ if (!Number.isSafeInteger(options.count) || options.count < 0) throw new Error(`${options.label} count must be a non-negative integer`);
52
+ if (!Number.isSafeInteger(options.maxConcurrency) || options.maxConcurrency < 1) throw new Error(`${options.label} maxConcurrency must be a positive integer`);
53
+ const controller = new AbortController();
54
+ const abortFromCaller = () => controller.abort(options.signal?.reason);
55
+ if (options.signal?.aborted) abortFromCaller();
56
+ else options.signal?.addEventListener("abort", abortFromCaller, { once: true });
57
+ const results = new Array(options.count);
58
+ let nextIndex = 0;
59
+ let failed = false;
60
+ let firstFailure;
61
+ const workers = Array.from({ length: Math.min(options.maxConcurrency, options.count) }, async () => {
62
+ while (!controller.signal.aborted) {
63
+ const index = nextIndex;
64
+ nextIndex += 1;
65
+ if (index >= options.count) return;
66
+ try {
67
+ results[index] = await options.map(index, controller.signal);
68
+ } catch (error) {
69
+ if (!failed) {
70
+ failed = true;
71
+ firstFailure = error;
72
+ controller.abort(error);
73
+ }
74
+ return;
75
+ }
76
+ }
77
+ });
78
+ try {
79
+ await Promise.all(workers);
80
+ if (options.signal?.aborted) throw abortError(options.signal, options.label);
81
+ if (failed) throw firstFailure;
82
+ if (controller.signal.aborted) throw abortError(controller.signal, options.label);
83
+ return results;
84
+ } finally {
85
+ options.signal?.removeEventListener("abort", abortFromCaller);
86
+ }
87
+ }
88
+ /**
89
+ * Map independent work with a fixed worker count while preserving input order.
90
+ * After the first rejection, no new items start; already-running work is allowed
91
+ * to settle before the returned promise rejects. Partial results are discarded.
92
+ */
93
+ async function mapConcurrent(items, concurrency, map) {
94
+ return mapConcurrentRange({
95
+ count: items.length,
96
+ maxConcurrency: concurrency,
97
+ label: "mapConcurrent",
98
+ map(index) {
99
+ return map(items[index], index);
100
+ }
101
+ });
102
+ }
103
+ function abortError(signal, label) {
104
+ return signal.reason instanceof Error ? signal.reason : /* @__PURE__ */ new Error(`${label} aborted`);
105
+ }
106
+ //#endregion
107
+ export { mapConcurrent as n, mapConcurrentRange as r, Mutex as t };
108
+
109
+ //# sourceMappingURL=concurrency-MUjT7VjM.js.map
@@ -0,0 +1 @@
1
+ {"version":3,"file":"concurrency-MUjT7VjM.js","names":[],"sources":["../src/concurrency.ts"],"sourcesContent":["/**\n * concurrency — small primitives the evolution loop needs.\n *\n * `Mutex` is a zero-dep async lock with FIFO fairness. The evolution loop\n * uses it to serialise checkout/build/commit sequences inside a single\n * pool slot, and to gate concurrent JSONL writers (see\n * `lockedJsonlReferenceReplayStore`).\n *\n * Deliberately minimal — no priority queue, no timeouts. If you need\n * those, swap to `async-mutex` at the call site.\n */\n\nexport class Mutex {\n private locked = false\n private readonly waiters: Array<() => void> = []\n\n async acquire(): Promise<() => void> {\n if (!this.locked) {\n this.locked = true\n return () => this.release()\n }\n return new Promise<() => void>((resolve) => {\n this.waiters.push(() => {\n resolve(() => this.release())\n })\n })\n }\n\n private release(): void {\n const next = this.waiters.shift()\n if (next) {\n next()\n } else {\n this.locked = false\n }\n }\n\n async runExclusive<T>(fn: () => Promise<T> | T): Promise<T> {\n const release = await this.acquire()\n try {\n return await fn()\n } finally {\n release()\n }\n }\n\n /** True iff someone holds the lock right now. Diagnostics only. */\n get isLocked(): boolean {\n return this.locked\n }\n\n /** Pending waiter count. Diagnostics only. */\n get pending(): number {\n return this.waiters.length\n }\n}\n\nexport interface MapConcurrentRangeOptions<R> {\n count: number\n maxConcurrency: number\n label: string\n signal?: AbortSignal\n map(index: number, signal: AbortSignal): Promise<R>\n}\n\n/** Map an integer range with bounded work, cancellation, and first-error cleanup. */\nexport async function mapConcurrentRange<R>(options: MapConcurrentRangeOptions<R>): Promise<R[]> {\n if (!Number.isSafeInteger(options.count) || options.count < 0) {\n throw new Error(`${options.label} count must be a non-negative integer`)\n }\n if (!Number.isSafeInteger(options.maxConcurrency) || options.maxConcurrency < 1) {\n throw new Error(`${options.label} maxConcurrency must be a positive integer`)\n }\n\n const controller = new AbortController()\n const abortFromCaller = () => controller.abort(options.signal?.reason)\n if (options.signal?.aborted) abortFromCaller()\n else options.signal?.addEventListener('abort', abortFromCaller, { once: true })\n\n const results = new Array<R>(options.count)\n let nextIndex = 0\n let failed = false\n let firstFailure: unknown\n const workers = Array.from(\n { length: Math.min(options.maxConcurrency, options.count) },\n async () => {\n while (!controller.signal.aborted) {\n const index = nextIndex\n nextIndex += 1\n if (index >= options.count) return\n try {\n results[index] = await options.map(index, controller.signal)\n } catch (error) {\n if (!failed) {\n failed = true\n firstFailure = error\n controller.abort(error)\n }\n return\n }\n }\n },\n )\n\n try {\n await Promise.all(workers)\n if (options.signal?.aborted) throw abortError(options.signal, options.label)\n if (failed) throw firstFailure\n if (controller.signal.aborted) throw abortError(controller.signal, options.label)\n return results\n } finally {\n options.signal?.removeEventListener('abort', abortFromCaller)\n }\n}\n\n/**\n * Map independent work with a fixed worker count while preserving input order.\n * After the first rejection, no new items start; already-running work is allowed\n * to settle before the returned promise rejects. Partial results are discarded.\n */\nexport async function mapConcurrent<T, R>(\n items: readonly T[],\n concurrency: number,\n map: (item: T, index: number) => Promise<R>,\n): Promise<R[]> {\n return mapConcurrentRange({\n count: items.length,\n maxConcurrency: concurrency,\n label: 'mapConcurrent',\n map(index) {\n return map(items[index]!, index)\n },\n })\n}\n\nfunction abortError(signal: AbortSignal, label: string): Error {\n return signal.reason instanceof Error ? signal.reason : new Error(`${label} aborted`)\n}\n"],"mappings":";;;;;;;;;;;;AAYA,IAAa,QAAb,MAAmB;CACjB,SAAiB;CACjB,UAA8C,CAAC;CAE/C,MAAM,UAA+B;EACnC,IAAI,CAAC,KAAK,QAAQ;GAChB,KAAK,SAAS;GACd,aAAa,KAAK,QAAQ;EAC5B;EACA,OAAO,IAAI,SAAqB,YAAY;GAC1C,KAAK,QAAQ,WAAW;IACtB,cAAc,KAAK,QAAQ,CAAC;GAC9B,CAAC;EACH,CAAC;CACH;CAEA,UAAwB;EACtB,MAAM,OAAO,KAAK,QAAQ,MAAM;EAChC,IAAI,MACF,KAAK;OAEL,KAAK,SAAS;CAElB;CAEA,MAAM,aAAgB,IAAsC;EAC1D,MAAM,UAAU,MAAM,KAAK,QAAQ;EACnC,IAAI;GACF,OAAO,MAAM,GAAG;EAClB,UAAU;GACR,QAAQ;EACV;CACF;;CAGA,IAAI,WAAoB;EACtB,OAAO,KAAK;CACd;;CAGA,IAAI,UAAkB;EACpB,OAAO,KAAK,QAAQ;CACtB;AACF;;AAWA,eAAsB,mBAAsB,SAAqD;CAC/F,IAAI,CAAC,OAAO,cAAc,QAAQ,KAAK,KAAK,QAAQ,QAAQ,GAC1D,MAAM,IAAI,MAAM,GAAG,QAAQ,MAAM,sCAAsC;CAEzE,IAAI,CAAC,OAAO,cAAc,QAAQ,cAAc,KAAK,QAAQ,iBAAiB,GAC5E,MAAM,IAAI,MAAM,GAAG,QAAQ,MAAM,2CAA2C;CAG9E,MAAM,aAAa,IAAI,gBAAgB;CACvC,MAAM,wBAAwB,WAAW,MAAM,QAAQ,QAAQ,MAAM;CACrE,IAAI,QAAQ,QAAQ,SAAS,gBAAgB;MACxC,QAAQ,QAAQ,iBAAiB,SAAS,iBAAiB,EAAE,MAAM,KAAK,CAAC;CAE9E,MAAM,UAAU,IAAI,MAAS,QAAQ,KAAK;CAC1C,IAAI,YAAY;CAChB,IAAI,SAAS;CACb,IAAI;CACJ,MAAM,UAAU,MAAM,KACpB,EAAE,QAAQ,KAAK,IAAI,QAAQ,gBAAgB,QAAQ,KAAK,EAAE,GAC1D,YAAY;EACV,OAAO,CAAC,WAAW,OAAO,SAAS;GACjC,MAAM,QAAQ;GACd,aAAa;GACb,IAAI,SAAS,QAAQ,OAAO;GAC5B,IAAI;IACF,QAAQ,SAAS,MAAM,QAAQ,IAAI,OAAO,WAAW,MAAM;GAC7D,SAAS,OAAO;IACd,IAAI,CAAC,QAAQ;KACX,SAAS;KACT,eAAe;KACf,WAAW,MAAM,KAAK;IACxB;IACA;GACF;EACF;CACF,CACF;CAEA,IAAI;EACF,MAAM,QAAQ,IAAI,OAAO;EACzB,IAAI,QAAQ,QAAQ,SAAS,MAAM,WAAW,QAAQ,QAAQ,QAAQ,KAAK;EAC3E,IAAI,QAAQ,MAAM;EAClB,IAAI,WAAW,OAAO,SAAS,MAAM,WAAW,WAAW,QAAQ,QAAQ,KAAK;EAChF,OAAO;CACT,UAAU;EACR,QAAQ,QAAQ,oBAAoB,SAAS,eAAe;CAC9D;AACF;;;;;;AAOA,eAAsB,cACpB,OACA,aACA,KACc;CACd,OAAO,mBAAmB;EACxB,OAAO,MAAM;EACb,gBAAgB;EAChB,OAAO;EACP,IAAI,OAAO;GACT,OAAO,IAAI,MAAM,QAAS,KAAK;EACjC;CACF,CAAC;AACH;AAEA,SAAS,WAAW,QAAqB,OAAsB;CAC7D,OAAO,OAAO,kBAAkB,QAAQ,OAAO,yBAAS,IAAI,MAAM,GAAG,MAAM,SAAS;AACtF"}
@@ -5,10 +5,10 @@ import { T as AnalystFinding, n as buildDefaultAnalystRegistry, t as DefaultAnal
5
5
  import { i as InMemoryOutcomeStore, n as FileSystemOutcomeStore, o as OutcomeStore, r as FileSystemOutcomeStoreOptions, t as DeploymentOutcome } from "../outcome-store-BYHIuO0e.js";
6
6
  import { _ as CodeAgentSessionExecutionReceipt, a as ParsedCodeAgentJsonl, b as CodeAgentSessionTerminalStatus, c as fromKimiCodeSession, d as fromPigraphSession, f as parseCodeAgentJsonl, g as CodeAgentSessionActionSurface, h as CodeAgentSessionActionStatus, i as CodeAgentSessionMetrics, l as fromOpenCodeSession, m as CodeAgentSessionActionKind, n as CodeAgentSessionIntakeOptions, o as fromClaudeCodeSession, p as CodeAgentSessionAction, r as CodeAgentSessionIntakeResult, s as fromCodexSession, t as CodeAgentSessionDiagnostic, u as fromPiSession, v as CodeAgentSessionObservation, x as observeCodeAgentSession, y as CodeAgentSessionSource } from "../code-agent-session-DqqgOJaz.js";
7
7
  import { C as JudgeDimension, H as SurfaceProposer, M as OptimizerConfig, R as Scenario, S as JudgeConfig, V as SessionScript, _ as GateDecision, a as CampaignResult, b as GenerationRecord, c as CampaignTraceWriter, d as DispatchContext, f as DispatchFn, g as GateContribution, h as GateContext, i as CampaignCostMeter, j as MutableSurface, k as LabeledScenarioStore, l as CodeSurface, m as GateCheckStatus, n as CampaignArtifactWriter, p as Gate, r as CampaignCellResult, t as CampaignAggregates, v as GateResult, w as JudgeScore, y as GenerationCandidate } from "../types-k9tZGKUg.js";
8
- import { A as RunEvalOptions, B as OptimizerModelBudget, Bt as ComparisonCost, C as RunImprovementLoopOptions, Cn as ReferenceEquivalenceJudgeResult, D as RunOptimizationOptions, Dn as LlmJudgeDimension, E as PremeasuredOptimizationBaseline, En as runReferenceEquivalenceJudge, F as GepaOptimizationMethodConfig, Ft as ExternalTextOptimizerContext, G as ObjectiveSource, Gt as OptimizationMethodProvenance, H as AxisVerdict, Ht as OptimizationMethodComparison, I as GepaOptimizationRecipe, It as ExternalTextOptimizerResult, J as PromotionPolicy, K as ParetoSignificanceGateOptions, Kt as OptimizationMethodResult, L as GepaRunnerCommand, Lt as ExternalOptimizationExample, M as GepaAdaptiveEngineRun, Mt as composeGate, N as GepaEngineOptions, Nt as externalTextOptimizationMethod, On as LlmJudgeOptions, P as GepaEngineRun, Pt as ExternalTextOptimizationMethodConfig, R as gepaOptimizationMethod, Rt as ExternalTextEvaluationResponse, Sn as ReferenceEquivalenceJudgeOptions, T as runImprovementLoop, Tn as createReferenceEquivalenceJudge, U as BuildEvidenceVectorOptions, Ut as OptimizationMethodInput, V as AxisEvidence, Vt as OptimizationMethod, W as EvidenceVector, X as paretoPolicy, Xt as OptimizationTokenUsage, Y as buildEvidenceVector, Yt as OptimizationPackageSource, Z as paretoSignificanceGate, Zt as compareOptimizationMethods, bn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, dt as DefaultProductionGateCheck, en as CampaignCellFailureReceipt, ft as DefaultProductionGateOptions, i as skillOptOptimizationMethod, in as RunCampaignOptions, j as runEval, kn as llmJudge, ln as fsCampaignStorage, lt as HeldOutGateOptions, mn as campaignSplitDigest, mt as defaultProductionGate, n as SkillOptRunnerCommand, on as runCampaign, ot as PowerPreflight, p as LoopProvenanceRecord, pt as DefaultProductionRewardHackingOptions, q as PromotionObjective, r as SkillOptTrainerConfig, sn as CampaignStorage, t as SkillOptOptimizationMethodConfig, un as inMemoryCampaignStorage, ut as heldOutGate, w as RunImprovementLoopResult, wn as ReferenceEquivalenceScenario, xn as ReferenceEquivalenceJudgeInput, yn as REFERENCE_EQUIVALENCE_INPUT_LIMITS, z as OpenAICompatibleOptimizerModel, zt as CompareOptimizationMethodsOptions } from "../skillopt-optimization-method-B7wX7XkF.js";
8
+ import { A as RunEvalOptions, B as OptimizerModelBudget, Bt as ComparisonCost, C as RunImprovementLoopOptions, Cn as ReferenceEquivalenceJudgeResult, D as RunOptimizationOptions, Dn as LlmJudgeDimension, E as PremeasuredOptimizationBaseline, En as runReferenceEquivalenceJudge, F as GepaOptimizationMethodConfig, Ft as ExternalTextOptimizerContext, G as ObjectiveSource, Gt as OptimizationMethodProvenance, H as AxisVerdict, Ht as OptimizationMethodComparison, I as GepaOptimizationRecipe, It as ExternalTextOptimizerResult, J as PromotionPolicy, K as ParetoSignificanceGateOptions, Kt as OptimizationMethodResult, L as GepaRunnerCommand, Lt as ExternalOptimizationExample, M as GepaAdaptiveEngineRun, Mt as composeGate, N as GepaEngineOptions, Nt as externalTextOptimizationMethod, On as LlmJudgeOptions, P as GepaEngineRun, Pt as ExternalTextOptimizationMethodConfig, R as gepaOptimizationMethod, Rt as ExternalTextEvaluationResponse, Sn as ReferenceEquivalenceJudgeOptions, T as runImprovementLoop, Tn as createReferenceEquivalenceJudge, U as BuildEvidenceVectorOptions, Ut as OptimizationMethodInput, V as AxisEvidence, Vt as OptimizationMethod, W as EvidenceVector, X as paretoPolicy, Xt as OptimizationTokenUsage, Y as buildEvidenceVector, Yt as OptimizationPackageSource, Z as paretoSignificanceGate, Zt as compareOptimizationMethods, bn as REFERENCE_EQUIVALENCE_JUDGE_VERSION, dt as DefaultProductionGateCheck, en as CampaignCellFailureReceipt, ft as DefaultProductionGateOptions, i as skillOptOptimizationMethod, in as RunCampaignOptions, j as runEval, kn as llmJudge, ln as fsCampaignStorage, lt as HeldOutGateOptions, mn as campaignSplitDigest, mt as defaultProductionGate, n as SkillOptRunnerCommand, on as runCampaign, ot as PowerPreflight, p as LoopProvenanceRecord, pt as DefaultProductionRewardHackingOptions, q as PromotionObjective, r as SkillOptTrainerConfig, sn as CampaignStorage, t as SkillOptOptimizationMethodConfig, un as inMemoryCampaignStorage, ut as heldOutGate, w as RunImprovementLoopResult, wn as ReferenceEquivalenceScenario, xn as ReferenceEquivalenceJudgeInput, yn as REFERENCE_EQUIVALENCE_INPUT_LIMITS, z as OpenAICompatibleOptimizerModel, zt as CompareOptimizationMethodsOptions } from "../skillopt-optimization-method-CWKVTnks.js";
9
9
  import { A as ScalarDistribution, C as InsightReport, D as OutcomeCorrelationInsight, E as LiftInsight, O as Recommendation, S as FailureClusterInsight, T as JudgeInsight, b as ExecutionInsight, c as EvalRunGenerationSnapshot, g as TraceSpanEvent, j as TokenUsageInsight, k as ReleaseSummary, n as HostedTenant, o as EvalRunCellScore, s as EvalRunEvent, v as CostProvenanceSummary, w as InterRaterInsight, x as FailureClassTally, y as ExecutionErrorOutcomeCell } from "../client-C97NMzqi.js";
10
10
  import { a as summarizeExecution, i as analyzeRuns, n as ExecutionReport, r as SummarizeExecutionOptions, t as AnalyzeRunsOptions } from "../analyze-runs-FsgYCinh.js";
11
- import { AgentCandidateBenchmarkCellRef, AgentCandidateBenchmarkSuiteInputs, AgentCandidateBenchmarkTask, AgentCandidateBenchmarkTaskMaterial, AgentCandidateBundle, AgentCandidateEvaluationPolicy, AgentCandidateExperiment, AgentCandidateExperimentMaterial, AgentCandidateExperimentMeasurement, AgentImprovementMeasuredComparison, CandidateExecutionEvidence } from "@tangle-network/agent-interface";
11
+ import { AgentCandidateBenchmarkCellRef, AgentCandidateBenchmarkSuiteInputs, AgentCandidateBenchmarkTask, AgentCandidateBenchmarkTaskMaterial, AgentCandidateBundle, AgentCandidateEvaluationPolicy, AgentCandidateExperiment, AgentCandidateExperimentMaterial, AgentCandidateExperimentMeasurement, AgentImprovementMeasuredComparison, AgentProfileImprovementExperiment, AgentProfileImprovementExperimentMaterial, AgentProfileImprovementMeasuredComparison, AgentProfileImprovementMeasurement, AgentProfileImprovementRunCell, AgentProfileImprovementRunReceipt, AgentProfileImprovementSuiteInputs, AgentProfileImprovementTask, AgentProfileImprovementTaskMaterial, CandidateExecutionEvidence, Sha256Digest } from "@tangle-network/agent-interface";
12
12
  //#region src/contract/self-improve.d.ts
13
13
  interface SelfImproveBudget {
14
14
  /** Hard spend cap across the full run. Each paid call reserves its enforced
@@ -406,6 +406,7 @@ interface CandidateExperimentExecutionInput {
406
406
  interface RunCandidateExperimentOptions {
407
407
  experiment: AgentCandidateExperiment;
408
408
  execute(input: CandidateExperimentExecutionInput): Promise<CandidateExecutionEvidence>;
409
+ /** Maximum number of simultaneous execute calls across both arms. */
409
410
  maxConcurrency?: number;
410
411
  signal?: AbortSignal;
411
412
  }
@@ -484,6 +485,59 @@ declare function verifyCandidateBenchmarkSuite(input: unknown): {
484
485
  digest: `sha256:${string}`;
485
486
  };
486
487
  //#endregion
488
+ //#region src/contract/profile-measured-comparison.d.ts
489
+ interface SealAgentProfileImprovementSuiteOptions {
490
+ splitDigest: Sha256Digest;
491
+ tasks: [AgentProfileImprovementTask, ...AgentProfileImprovementTask[]];
492
+ reps: number;
493
+ seeds: [number, ...number[]];
494
+ }
495
+ interface AgentProfileImprovementExperimentExecutionInput {
496
+ experiment: AgentProfileImprovementExperiment;
497
+ arm: 'baseline' | 'candidate';
498
+ stateDigest: Sha256Digest;
499
+ task: AgentProfileImprovementTask;
500
+ runCell: AgentProfileImprovementRunCell;
501
+ seed: number;
502
+ signal?: AbortSignal;
503
+ }
504
+ interface RunAgentProfileImprovementExperimentOptions {
505
+ experiment: AgentProfileImprovementExperiment;
506
+ execute(input: AgentProfileImprovementExperimentExecutionInput): Promise<AgentProfileImprovementRunReceipt>;
507
+ /** Maximum number of simultaneous execute calls across both arms. */
508
+ maxConcurrency?: number;
509
+ signal?: AbortSignal;
510
+ }
511
+ interface CompareAgentProfileImprovementExperimentOptions {
512
+ experiment: AgentProfileImprovementExperiment;
513
+ measurements: AgentProfileImprovementMeasurement[];
514
+ runId: string;
515
+ candidate?: AgentProfileImprovementMeasuredComparison['candidate'];
516
+ generationsExplored?: number;
517
+ searchDurationMs?: number;
518
+ searchCostUsd?: number;
519
+ metadata?: AgentProfileImprovementMeasuredComparison['metadata'];
520
+ }
521
+ /** Content-address one held-out profile task before either state can execute it. */
522
+ declare function sealAgentProfileImprovementTask(material: AgentProfileImprovementTaskMaterial): AgentProfileImprovementTask;
523
+ /** Freeze profile task order, repetitions, seeds, and the held-out split. */
524
+ declare function sealAgentProfileImprovementSuite(options: SealAgentProfileImprovementSuiteOptions): AgentProfileImprovementSuiteInputs;
525
+ /** Freeze the two host-owned profile states and their exact held-out work. */
526
+ declare function sealAgentProfileImprovementExperiment(material: AgentProfileImprovementExperimentMaterial): AgentProfileImprovementExperiment;
527
+ declare function verifyAgentProfileImprovementTask(input: unknown): AgentProfileImprovementTask;
528
+ declare function verifyAgentProfileImprovementSuiteInputs(input: unknown): AgentProfileImprovementSuiteInputs;
529
+ declare function verifyAgentProfileImprovementExperiment(input: unknown): AgentProfileImprovementExperiment;
530
+ /**
531
+ * Execute each signed profile cell through the host's one exact-state executor.
532
+ * Eval owns only the cell schedule and receipt checks; the host resolves each
533
+ * state digest and captures its own run, billing, trace, and grader evidence.
534
+ */
535
+ declare function runAgentProfileImprovementExperiment(options: RunAgentProfileImprovementExperimentOptions): Promise<AgentProfileImprovementMeasurement[]>;
536
+ /** Build the only publishable profile comparison from complete host receipts. */
537
+ declare function measuredComparisonFromAgentProfileImprovementExperiment(options: CompareAgentProfileImprovementExperimentOptions): AgentProfileImprovementMeasuredComparison;
538
+ /** Recompute a profile comparison from the exact sealed experiment and receipts. */
539
+ declare function verifyAgentProfileImprovementExperimentComparison(input: unknown): AgentProfileImprovementMeasuredComparison;
540
+ //#endregion
487
541
  //#region src/contract/intake/run-record-dir.d.ts
488
542
  /** A record that failed boundary validation, with enough context to fix it. */
489
543
  interface RunRecordRejection {
@@ -845,5 +899,5 @@ interface FromOtelSpansOptions {
845
899
  }
846
900
  declare function fromOtelSpans(opts: FromOtelSpansOptions): RunRecord[];
847
901
  //#endregion
848
- export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentTraceContributor, type AgentTraceContributorType, type AgentTraceConversation, type AgentTraceFile, type AgentTraceIndex, type AgentTraceRange, type AgentTraceRecord, type AnalystFinding, type AnalyzeRunsOptions, type AuthoringProvenance, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type CampaignAggregates, type CampaignArtifactWriter, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type ChatClient, type CodeAgentSessionAction, type CodeAgentSessionActionKind, type CodeAgentSessionActionStatus, type CodeAgentSessionActionSurface, type CodeAgentSessionDiagnostic, type CodeAgentSessionExecutionReceipt, type CodeAgentSessionIntakeOptions, type CodeAgentSessionIntakeResult, type CodeAgentSessionMetrics, type CodeAgentSessionObservation, type CodeAgentSessionSource, type CodeAgentSessionTerminalStatus, type CodeSurface, type CompareCandidateExperimentOptions, type CompareOptimizationMethodsOptions, type ComparisonCost, type CostLedgerHandle, type CostProvenanceSummary, type CreateChatClientOpts, type DefaultAnalystRegistryOptions, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DefineAgentEvalOptions, type DefinedAgentEval, type DeploymentOutcome, type DispatchFn as Dispatch, type DispatchContext, type EvalCellScoreDelta, type EvalDimensionDelta, type EvalGenerationDiff, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvalRunDiff, type EvaluatePairedMeasurementsOptions, type EvidenceVector, type ExecutionErrorOutcomeCell, type ExecutionInsight, type ExecutionReport, type ExternalOptimizationExample, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, type FailureClassTally, type FailureClusterInsight, type FeedbackTableMeta, type FeedbackTableRow, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, type FromFeedbackTableOptions, type FromFeedbackTableResult, type FromOtelSpansOptions, type FromRunRecordDirOptions, type FromRunRecordDirResult, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaEngineOptions, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaRunnerCommand, type HeldOutGateOptions, type HostedTenant, InMemoryOutcomeStore, type InsightReport, type InterRaterInsight, type JudgeConfig, type JudgeDimension, type JudgeInsight, type JudgeScore, type LiftInsight, type LlmJudgeDimension, type LlmJudgeOptions, type MutableSurface, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationPackageSource, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type OutcomeCorrelationInsight, type OutcomeStore, type PairedMeasurement, type PairedMeasurementAdapter, type PairedMeasurementEvaluation, type ParetoSignificanceGateOptions, type ParsedCodeAgentJsonl, type PartitionByAuthoringModelResult, type PromotionObjective, type PromotionPolicy, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveResult, SelfImproveRunError, type SessionScript, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SummarizeExecutionOptions, type SurfaceProposer, type TokenUsageInsight, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, diffGenerations, diffRunBaselineToWinner, diffRuns, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fromPigraphSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, partitionRunsByAuthoringModel, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, summarizeExecution, verifyCandidateBenchmarkSuite, verifyCandidateBenchmarkSuiteInputs, verifyCandidateBenchmarkTask, verifyCandidateExperiment, verifyCandidateExperimentComparison };
902
+ export { type AgentEvalAgent, type AgentEvalEvaluateOptions, type AgentEvalImproveOptions, type AgentProfileImprovementExperimentExecutionInput, type AgentTraceContributor, type AgentTraceContributorType, type AgentTraceConversation, type AgentTraceFile, type AgentTraceIndex, type AgentTraceRange, type AgentTraceRecord, type AnalystFinding, type AnalyzeRunsOptions, type AuthoringProvenance, type AxisEvidence, type AxisVerdict, type BuildEvidenceVectorOptions, type CampaignAggregates, type CampaignArtifactWriter, type CampaignCellFailureReceipt, type CampaignCellResult, type CampaignCostMeter, type CampaignResult, type CampaignStorage, type CampaignTraceWriter, type CandidateExperimentExecutionInput, type ChatClient, type CodeAgentSessionAction, type CodeAgentSessionActionKind, type CodeAgentSessionActionStatus, type CodeAgentSessionActionSurface, type CodeAgentSessionDiagnostic, type CodeAgentSessionExecutionReceipt, type CodeAgentSessionIntakeOptions, type CodeAgentSessionIntakeResult, type CodeAgentSessionMetrics, type CodeAgentSessionObservation, type CodeAgentSessionSource, type CodeAgentSessionTerminalStatus, type CodeSurface, type CompareAgentProfileImprovementExperimentOptions, type CompareCandidateExperimentOptions, type CompareOptimizationMethodsOptions, type ComparisonCost, type CostLedgerHandle, type CostProvenanceSummary, type CreateChatClientOpts, type DefaultAnalystRegistryOptions, type DefaultProductionGateCheck, type DefaultProductionGateOptions, type DefaultProductionRewardHackingOptions, type DefineAgentEvalOptions, type DefinedAgentEval, type DeploymentOutcome, type DispatchFn as Dispatch, type DispatchContext, type EvalCellScoreDelta, type EvalDimensionDelta, type EvalGenerationDiff, type EvalReportingSuiteInput, type EvalReportingSuiteOptions, type EvalReportingSuiteResult, type EvalRunDiff, type EvaluatePairedMeasurementsOptions, type EvidenceVector, type ExecutionErrorOutcomeCell, type ExecutionInsight, type ExecutionReport, type ExternalOptimizationExample, type ExternalTextEvaluationResponse, type ExternalTextOptimizationMethodConfig, type ExternalTextOptimizerContext, type ExternalTextOptimizerResult, type FailureClassTally, type FailureClusterInsight, type FeedbackTableMeta, type FeedbackTableRow, FileSystemOutcomeStore, type FileSystemOutcomeStoreOptions, type FromFeedbackTableOptions, type FromFeedbackTableResult, type FromOtelSpansOptions, type FromRunRecordDirOptions, type FromRunRecordDirResult, type Gate, type GateCheckStatus, type GateContext, type GateContribution, type GateDecision, type GateResult, type GenerationCandidate, type GenerationRecord, type GepaAdaptiveEngineRun, type GepaEngineOptions, type GepaEngineRun, type GepaOptimizationMethodConfig, type GepaOptimizationRecipe, type GepaRunnerCommand, type HeldOutGateOptions, type HostedTenant, InMemoryOutcomeStore, type InsightReport, type InterRaterInsight, type JudgeConfig, type JudgeDimension, type JudgeInsight, type JudgeScore, type LiftInsight, type LlmJudgeDimension, type LlmJudgeOptions, type MutableSurface, type ObjectiveSource, type OpenAICompatibleOptimizerModel, type OptimizationMethod, type OptimizationMethodComparison, type OptimizationMethodInput, type OptimizationMethodProvenance, type OptimizationMethodResult, type OptimizationPackageSource, type OptimizationTokenUsage, type OptimizerConfig, type OptimizerModelBudget, type OutcomeCorrelationInsight, type OutcomeStore, type PairedMeasurement, type PairedMeasurementAdapter, type PairedMeasurementEvaluation, type ParetoSignificanceGateOptions, type ParsedCodeAgentJsonl, type PartitionByAuthoringModelResult, type PromotionObjective, type PromotionPolicy, REFERENCE_EQUIVALENCE_INPUT_LIMITS, REFERENCE_EQUIVALENCE_JUDGE_VERSION, type Recommendation, type ReferenceEquivalenceJudgeInput, type ReferenceEquivalenceJudgeOptions, type ReferenceEquivalenceJudgeResult, type ReferenceEquivalenceScenario, type ReleaseSummary, type RunAgentProfileImprovementExperimentOptions, type RunCampaignOptions, type RunCandidateExperimentOptions, type RunEvalOptions, type RunImprovementLoopOptions, type RunImprovementLoopResult, type RunRecordRejection, type ScalarDistribution, type Scenario, type SealAgentProfileImprovementSuiteOptions, type SealCandidateBenchmarkSuiteOptions, type SelfImproveBudget, type SelfImproveOptions, type SelfImproveProgressEvent, type SelfImproveResult, SelfImproveRunError, type SessionScript, type SkillOptOptimizationMethodConfig, type SkillOptRunnerCommand, type SkillOptTrainerConfig, type SummarizeExecutionOptions, type SurfaceProposer, type TokenUsageInsight, analyzeRuns, buildDefaultAnalystRegistry, buildEvidenceVector, campaignSplitDigest, compareOptimizationMethods, composeGate, createChatClient, createReferenceEquivalenceJudge, defaultProductionGate, defineAgentEval, diffGenerations, diffRunBaselineToWinner, diffRuns, evalReportingSuite, evaluatePairedMeasurements, externalTextOptimizationMethod, fromClaudeCodeSession, fromCodexSession, fromFeedbackTable, fromKimiCodeSession, fromOpenCodeSession, fromOtelSpans, fromPiSession, fromPigraphSession, fromRunRecordDir, fsCampaignStorage, gepaOptimizationMethod, heldOutGate, inMemoryCampaignStorage, llmJudge, measuredComparisonFromAgentProfileImprovementExperiment, measuredComparisonFromCandidateExperiment, observeCodeAgentSession, paretoPolicy, paretoSignificanceGate, parseAgentTrace, parseCodeAgentJsonl, partitionRunsByAuthoringModel, runAgentProfileImprovementExperiment, runCampaign, runCandidateExperiment, runEval, runImprovementLoop, runReferenceEquivalenceJudge, sealAgentProfileImprovementExperiment, sealAgentProfileImprovementSuite, sealAgentProfileImprovementTask, sealCandidateBenchmarkSuite, sealCandidateBenchmarkTask, sealCandidateExperiment, selfImprove, skillOptOptimizationMethod, summarizeExecution, verifyAgentProfileImprovementExperiment, verifyAgentProfileImprovementExperimentComparison, verifyAgentProfileImprovementSuiteInputs, verifyAgentProfileImprovementTask, verifyCandidateBenchmarkSuite, verifyCandidateBenchmarkSuiteInputs, verifyCandidateBenchmarkTask, verifyCandidateExperiment, verifyCandidateExperimentComparison };
849
903
  //# sourceMappingURL=index.d.ts.map
@@ -1 +1 @@
1
- {"version":3,"file":"index.d.ts","names":[],"sources":["../../src/contract/self-improve.ts","../../src/contract/define-agent-eval.ts","../../src/contract/measured-comparison.ts","../../src/contract/intake/run-record-dir.ts","../../src/contract/eval-reporting-suite.ts","../../src/contract/diff.ts","../../src/contract/intake/agent-trace.ts","../../src/contract/intake/feedback-table.ts","../../src/contract/intake/otel-spans.ts"],"mappings":";;;;;;;;;;;;UA8DiB;;;EAGf;;;;EAIA;;EAEA;;EAEA;;;EAGA;;;EAGA;;;;EAIA;;EAEA,mBAAmB;;;;;;;;;EASnB;;EAEA;;;;;EAKA;;KAGU;EACN;EAA0B;;EAC1B;EAA4B;EAAuB;;EACnD;EAA4B;EAAe;;EAC3C;EAA8B;EAAe;EAAuB;;EAGpE;EAAsB;EAAkB;;EACxC;EAAyB;EAAW;EAAY;EAAa;;UAElD,mBAAmB,kBAAkB,UAAU;;;;;;;;;;;;;;EAc9D,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;;EAIA,WAAW;;;EAIX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;EAaT,sBAAsB,gCAAgC,WAAW;;;;;EAMjE,WAAW;;;;;;EAOX,SAAS,mBAAmB,WAAW;;;EAIvC,qBAAqB;;;EAIrB,OAAO,KAAK,WAAW;;;;;;;;EASvB,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;;;EAKjF,UAAU;;;;EAKV;;;EAIA,gBAAgB,QAAQ;;;;;EAMxB,iBAAiB;IACf,UAAU;IACV;IACA;;;;EAKF;;;EAIA,cAAc,OAAO;;;;EAKrB;EACA;EACA;;;;;;;;;;;;;;EAeA,eAAe;;;EAIf,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD;;;;;;;EAQA,mBAAmB,uBAAuB,WAAW;;UAGtC,kBAAkB,kBAAkB,UAAU;;;;EAI7D;IACE;IACA,aAAa;;;;;EAKf;IACE;IACA,aAAa;IACb,SAAS;;;IAGT;;;;IAIA;;;;;;EAMF;;;EAGA;;;;EAIA,YAAY;;EAEZ;;;EAGA;;EAEA;;EAEA;;EAEA,MAAM;;;EAGN,UAAU;;EAEV;IACE;IACA,MAAM;IACN;IACA,aAAa;;;;;;;;EAQf,SAAS;;;;;EAKT,QAAQ;;;;;;EAMR,KAAK,yBAAyB,WAAW;;;cAI9B,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA+LhB,YAAY,kBAAkB,UAAU,WAC5D,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KCtiB5B,eAAe,kBAAkB,UAAU,cACrD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;KAED,uBAAuB,kBAAkB,UAAU,aAAa,mBAC1E,WACA;UAGe,yBAAyB,kBAAkB,UAAU,mBAC5D,KACN,eAAe,WAAW;;EAI5B,YAAY;;EAEZ,UAAU;;EAEV,QAAQ,eAAe,WAAW;;EAElC,QAAQ,YAAY,WAAW;;EAE/B,SAAS,YAAY,WAAW;;EAEhC;;KAGU,wBAAwB,kBAAkB,UAAU,aAAa,KAC3E,QAAQ,mBAAmB,WAAW;EAGtC,SAAS,QAAQ;EACjB,eAAe,QAAQ;;UAGR,iBAAiB,kBAAkB,UAAU;;WAEnD,oBAAoB;;WAEpB,iBAAiB;;;;;EAK1B,SACE,OAAO,yBAAyB,WAAW,aAC1C,QAAQ,eAAe,WAAW;;;;;;;EAOrC,QACE,OAAO,wBAAwB,WAAW,aACzC,QAAQ,kBAAkB,WAAW;;;;;;;;;iBAU1B,gBAAgB,kBAAkB,UAAU,WAC1D,UAAU,uBAAuB,WAAW,aAC3C,iBAAiB,WAAW;;;UC7Dd;EACf,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,QAAQ;EACR,MAAM;EACN,eAAe;EACf;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QAAQ,OAAO,oCAAoC,QAAQ;EAC3D;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,cAAc;EACd;EACA,YAAY;EACZ;EACA;EACA;EACA,WAAW;;;UAII,kBAAkB;EACjC;EACA,UAAU;EACV,WAAW;;;UAII,yBAAyB;EACxC,MAAM,KAAK;EACX,WAAW,KAAK;IAAkB;IAAc;;EAChD,QAAQ,KAAK;EACb,UAAU,KAAK;EACf,UAAU,KAAK;EACf,OAAO,KAAK;;UAGG,kCAAkC;EACjD,uBAAuB,kBAAkB;EACzC,QAAQ;EACR,SAAS,yBAAyB;;EAElC;;EAEA;;;KAIU,8BAA8B,KACxC;EAGA;EACA;EACA;;;iBAIc,2BACd,UAAU,sCACT;;iBAQa,4BACd,SAAS,qCACR;;iBAiBa,wBACd,UAAU,mCACT;iBAQa,0BAA0B,iBAAiB;;iBAarC,uBACpB,SAAS,gCACR,QAAQ;;;;;;;;iBAiEK,2BAA2B,MACzC,SAAS,kCAAkC,QAC1C;;iBA+Na,0CACd,SAAS,oCACR;;iBAoEa,oCACd,iBACC;iBAwCa,6BAA6B,iBAAiB;iBAM9C,oCACd,iBACC;iBAkBa,8BAA8B;;;;;;;;;;;UCljB7B;;EAEf;;EAEA;;EAEA;;UAGe;;;;;;EAMf;;;;;;;EAOA,WAAW;;;;;;EAMX;;UAGe;;EAEf,MAAM;;EAEN,UAAU;;EAEV;;;;;;;;;;iBAkBoB,iBACpB,cACA,UAAS,0BACR,QAAQ;;;;;KC7CC,0BAA0B;UAErB;;;;;EAKf,UAAU,KAAK;;EAEf,OAAO;;;;;;;;;;EAUP;;;;UAKe;;;EAGf,QAAQ;;EAER;;IAEE;;IAEA;;;IAGA;;IAEA;;;IAGA,UAAU;;;EAGZ;;;;;;;iBAUoB,mBACpB,OAAO,yBACP,UAAS,4BACR,QAAQ;;;;;;UC5DM;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;;;EAGA,YAAY,eAAe,eAAe;;;;UAK3B;EACf;EACA;EACA;EACA;EACA;;EAEA,SAAS;;EAET,SAAS;;EAET,OAAO;;EAEP;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;;;UAMe;EACf;EACA;EACA;EACA;EACA,oBAAoB;EACpB,mBAAmB;EACnB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;EAEA,cAAc;;;;EAId,aAAa;;;;;;;;;iBAoDC,gBACd,QAAQ,2BACR,OAAO,4BACN;;;;;;iBAqEa,SAAS,QAAQ,cAAc,OAAO,eAAe;;;;;;;iBAsCrD,wBAAwB,KAAK,eAAe;;;KC1OhD;UAEK;EACf,MAAM;;EAEN;;UAGe;EACf;EACA;EACA;;;EAGA,cAAc;;UAGC;EACf;EACA,cAAc;EACd,QAAQ;;UAGO;EACf;EACA,eAAe;;UAGA;EACf;EACA;EACA;EACA;IAAQ;IAAc;;EACtB;IAAS;IAAe;;EACxB,OAAO;;;;UAOQ;EACf;;EAEA;;EAEA;EACA;EACA;;EAEA;;EAEA;;KAGU,kBAAkB,YAAY;;;;;;iBAW1B,gBAAgB,SAAS,qBAAqB;UAmE7C;;;;EAIf,SAAS,YAAY;;;EAGrB,cAAc;;;;;;;;iBASA,8BACd,MAAM,aACN,OAAO,kBACN;;;UC/Jc;;;EAGf;;EAEA;;;EAGA;;;EAGA,WAAW;;UAGI;EACf;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,WAAW;;;EAGX,SAAS;;UAGM;;EAEf,SAAS;;;EAGT,OAAO;;;;EAIP;IAAU;IAAa;;;;;EAIvB;;UAGe;EACf,MAAM;;;EAGN,aAAa;IAAQ;IAAe;IAAe;;;iBAGrC,kBAAkB,MAAM,2BAA2B;;;UCvBlD;EACf,OAAO;;EAEP,eAAe;;EAEf;;;;;;EAMA,eAAe,eAAe,gBAAgB;;iBAGhC,cAAc,MAAM,uBAAuB"}
1
+ {"version":3,"file":"index.d.ts","names":[],"sources":["../../src/contract/self-improve.ts","../../src/contract/define-agent-eval.ts","../../src/contract/measured-comparison.ts","../../src/contract/profile-measured-comparison.ts","../../src/contract/intake/run-record-dir.ts","../../src/contract/eval-reporting-suite.ts","../../src/contract/diff.ts","../../src/contract/intake/agent-trace.ts","../../src/contract/intake/feedback-table.ts","../../src/contract/intake/otel-spans.ts"],"mappings":";;;;;;;;;;;;UA8DiB;;;EAGf;;;;EAIA;;EAEA;;EAEA;;;EAGA;;;EAGA;;;;EAIA;;EAEA,mBAAmB;;;;;;;;;EASnB;;EAEA;;;;;EAKA;;KAGU;EACN;EAA0B;;EAC1B;EAA4B;EAAuB;;EACnD;EAA4B;EAAe;;EAC3C;EAA8B;EAAe;EAAuB;;EAGpE;EAAsB;EAAkB;;EACxC;EAAyB;EAAW;EAAY;EAAa;;UAElD,mBAAmB,kBAAkB,UAAU;;;;;;;;;;;;;;EAc9D,QAAQ,SAAS,gBAAgB,UAAU,WAAW,KAAK,oBAAoB,QAAQ;;;;;;;EAQvF;;;EAIA,WAAW;;;EAIX,OAAO,YAAY,WAAW;;;EAI9B,iBAAiB;;EAGjB,SAAS;;;;;;;;;;;;EAaT,sBAAsB,gCAAgC,WAAW;;;;;EAMjE,WAAW;;;;;;EAOX,SAAS,mBAAmB,WAAW;;;EAIvC,qBAAqB;;;EAIrB,OAAO,KAAK,WAAW;;;;;;;;EASvB,cAAc,eAAe,gBAAgB,iBAAiB,mBAAmB;;;;EAKjF,UAAU;;;;EAKV;;;EAIA,gBAAgB,QAAQ;;;;;EAMxB,iBAAiB;IACf,UAAU;IACV;IACA;;;;EAKF;;;EAIA,cAAc,OAAO;;;;EAKrB;EACA;EACA;;;;;;;;;;;;;;EAeA,eAAe;;;EAIf,eAAe;;;;EAKf,eAAe;;EAGf;;;;;;;;EASA;;;;;;;;;EAUA,oBAAoB,uBAAuB,WAAW;;;EAItD;;;;;;;EAQA,mBAAmB,uBAAuB,WAAW;;UAGtC,kBAAkB,kBAAkB,UAAU;;;;EAI7D;IACE;IACA,aAAa;;;;;EAKf;IACE;IACA,aAAa;IACb,SAAS;;;IAGT;;;;IAIA;;;;;;EAMF;;;EAGA;;;;EAIA,YAAY;;EAEZ;;;EAGA;;EAEA;;EAEA;;EAEA,MAAM;;;EAGN,UAAU;;EAEV;IACE;IACA,MAAM;IACN;IACA,aAAa;;;;;;;;EAQf,SAAS;;;;;EAKT,QAAQ;;;;;;EAMR,KAAK,yBAAyB,WAAW;;;cAI9B,4BAA4B;WAC9B,MAAM;WACN,UAAU;EAEnB,YAAY,gBAAgB,QAAQ;;;;;;;;;;;;;;;;;;;;;;;;;;;;iBA+LhB,YAAY,kBAAkB,UAAU,WAC5D,MAAM,mBAAmB,WAAW,aACnC,QAAQ,kBAAkB,WAAW;;;KCtiB5B,eAAe,kBAAkB,UAAU,cACrD,SAAS,gBACT,UAAU,WACV,KAAK,oBACF,QAAQ;KAED,uBAAuB,kBAAkB,UAAU,aAAa,mBAC1E,WACA;UAGe,yBAAyB,kBAAkB,UAAU,mBAC5D,KACN,eAAe,WAAW;;EAI5B,YAAY;;EAEZ,UAAU;;EAEV,QAAQ,eAAe,WAAW;;EAElC,QAAQ,YAAY,WAAW;;EAE/B,SAAS,YAAY,WAAW;;EAEhC;;KAGU,wBAAwB,kBAAkB,UAAU,aAAa,KAC3E,QAAQ,mBAAmB,WAAW;EAGtC,SAAS,QAAQ;EACjB,eAAe,QAAQ;;UAGR,iBAAiB,kBAAkB,UAAU;;WAEnD,oBAAoB;;WAEpB,iBAAiB;;;;;EAK1B,SACE,OAAO,yBAAyB,WAAW,aAC1C,QAAQ,eAAe,WAAW;;;;;;;EAOrC,QACE,OAAO,wBAAwB,WAAW,aACzC,QAAQ,kBAAkB,WAAW;;;;;;;;;iBAU1B,gBAAgB,kBAAkB,UAAU,WAC1D,UAAU,uBAAuB,WAAW,aAC3C,iBAAiB,WAAW;;;UC7Dd;EACf,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,QAAQ;EACR,MAAM;EACN,eAAe;EACf;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QAAQ,OAAO,oCAAoC,QAAQ;;EAE3D;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,cAAc;EACd;EACA,YAAY;EACZ;EACA;EACA;EACA,WAAW;;;UAII,kBAAkB;EACjC;EACA,UAAU;EACV,WAAW;;;UAII,yBAAyB;EACxC,MAAM,KAAK;EACX,WAAW,KAAK;IAAkB;IAAc;;EAChD,QAAQ,KAAK;EACb,UAAU,KAAK;EACf,UAAU,KAAK;EACf,OAAO,KAAK;;UAGG,kCAAkC;EACjD,uBAAuB,kBAAkB;EACzC,QAAQ;EACR,SAAS,yBAAyB;;EAElC;;EAEA;;;KAIU,8BAA8B,KACxC;EAGA;EACA;EACA;;;iBAIc,2BACd,UAAU,sCACT;;iBAQa,4BACd,SAAS,qCACR;;iBAiBa,wBACd,UAAU,mCACT;iBAQa,0BAA0B,iBAAiB;;iBAarC,uBACpB,SAAS,gCACR,QAAQ;;;;;;;;iBA0CK,2BAA2B,MACzC,SAAS,kCAAkC,QAC1C;;iBA2Na,0CACd,SAAS,oCACR;;iBAoEa,oCACd,iBACC;iBAwCa,6BAA6B,iBAAiB;iBAM9C,oCACd,iBACC;iBAkBa,8BAA8B;;;;;;;;;;UCnhB7B;EACf,aAAa;EACb,QAAQ,gCAAgC;EACxC;EACA;;UAGe;EACf,YAAY;EACZ;EACA,aAAa;EACb,MAAM;EACN,SAAS;EACT;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,QACE,OAAO,kDACN,QAAQ;;EAEX;EACA,SAAS;;UAGM;EACf,YAAY;EACZ,cAAc;EACd;EACA,YAAY;EACZ;EACA;EACA;EACA,WAAW;;;iBAIG,gCACd,UAAU,sCACT;;iBAQa,iCACd,SAAS,0CACR;;iBAqBa,sCACd,UAAU,4CACT;iBAOa,kCAAkC,iBAAiB;iBAInD,yCACd,iBACC;iBAIa,wCACd,iBACC;;;;;;iBASmB,qCACpB,SAAS,8CACR,QAAQ;;iBAmBK,wDACd,SAAS,kDACR;;iBAoEa,kDACd,iBACC;;;;UCxMc;;EAEf;;EAEA;;EAEA;;UAGe;;;;;;EAMf;;;;;;;EAOA,WAAW;;;;;;EAMX;;UAGe;;EAEf,MAAM;;EAEN,UAAU;;EAEV;;;;;;;;;;iBAkBoB,iBACpB,cACA,UAAS,0BACR,QAAQ;;;;;KC7CC,0BAA0B;UAErB;;;;;EAKf,UAAU,KAAK;;EAEf,OAAO;;;;;;;;;;EAUP;;;;UAKe;;;EAGf,QAAQ;;EAER;;IAEE;;IAEA;;;IAGA;;IAEA;;;IAGA,UAAU;;;EAGZ;;;;;;;iBAUoB,mBACpB,OAAO,yBACP,UAAS,4BACR,QAAQ;;;;;;UC5DM;EACf;EACA;EACA;;;UAIe;EACf;EACA;EACA;EACA;EACA;;;EAGA,YAAY,eAAe,eAAe;;;;UAK3B;EACf;EACA;EACA;EACA;EACA;;EAEA,SAAS;;EAET,SAAS;;EAET,OAAO;;EAEP;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;;;;UAMe;EACf;EACA;EACA;EACA;EACA,oBAAoB;EACpB,mBAAmB;EACnB;EACA;EACA;EACA;EACA;EACA;EACA;EACA;EACA;;EAEA,cAAc;;;;EAId,aAAa;;;;;;;;;iBAoDC,gBACd,QAAQ,2BACR,OAAO,4BACN;;;;;;iBAqEa,SAAS,QAAQ,cAAc,OAAO,eAAe;;;;;;;iBAsCrD,wBAAwB,KAAK,eAAe;;;KC1OhD;UAEK;EACf,MAAM;;EAEN;;UAGe;EACf;EACA;EACA;;;EAGA,cAAc;;UAGC;EACf;EACA,cAAc;EACd,QAAQ;;UAGO;EACf;EACA,eAAe;;UAGA;EACf;EACA;EACA;EACA;IAAQ;IAAc;;EACtB;IAAS;IAAe;;EACxB,OAAO;;;;UAOQ;EACf;;EAEA;;EAEA;EACA;EACA;;EAEA;;EAEA;;KAGU,kBAAkB,YAAY;;;;;;iBAW1B,gBAAgB,SAAS,qBAAqB;UAmE7C;;;;EAIf,SAAS,YAAY;;;EAGrB,cAAc;;;;;;;;iBASA,8BACd,MAAM,aACN,OAAO,kBACN;;;UC/Jc;;;EAGf;;EAEA;;;EAGA;;;EAGA,WAAW;;UAGI;EACf;;;EAGA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;EAEA;;;EAGA,WAAW;;;EAGX,SAAS;;UAGM;;EAEf,SAAS;;;EAGT,OAAO;;;;EAIP;IAAU;IAAa;;;;;EAIvB;;UAGe;EACf,MAAM;;;EAGN,aAAa;IAAQ;IAAe;IAAe;;;iBAGrC,kBAAkB,MAAM,2BAA2B;;;UCvBlD;EACf,OAAO;;EAEP,eAAe;;EAEf;;;;;;EAMA,eAAe,eAAe,gBAAgB;;iBAGhC,cAAc,MAAM,uBAAuB"}