@tangle-network/agent-eval 0.115.3 → 0.116.0

package/dist/chunk-MPHTT5HE.js

@@ -0,0 +1,74 @@
+// src/concurrency.ts
+var Mutex = class {
+  locked = false;
+  waiters = [];
+  async acquire() {
+    if (!this.locked) {
+      this.locked = true;
+      return () => this.release();
+    }
+    return new Promise((resolve) => {
+      this.waiters.push(() => {
+        resolve(() => this.release());
+      });
+    });
+  }
+  release() {
+    const next = this.waiters.shift();
+    if (next) {
+      next();
+    } else {
+      this.locked = false;
+    }
+  }
+  async runExclusive(fn) {
+    const release = await this.acquire();
+    try {
+      return await fn();
+    } finally {
+      release();
+    }
+  }
+  /** True iff someone holds the lock right now. Diagnostics only. */
+  get isLocked() {
+    return this.locked;
+  }
+  /** Pending waiter count. Diagnostics only. */
+  get pending() {
+    return this.waiters.length;
+  }
+};
+
+// src/run-score.ts
+var DEFAULT_RUN_SCORE_WEIGHTS = {
+  success: 4,
+  goalProgress: 2,
+  repoGroundedness: 1.5,
+  driftPenalty: -1.5,
+  toolUseQuality: 1,
+  patchQuality: 1.25,
+  testReality: 1.5,
+  finalGate: 3,
+  reviewerBlockers: -2,
+  costUsd: -0.2,
+  wallSeconds: -0.1
+};
+function aggregateRunScore(score, weights = {}) {
+  const w = { ...DEFAULT_RUN_SCORE_WEIGHTS, ...weights };
+  return w.success * clamp01(score.success) + w.goalProgress * clamp01(score.goalProgress) + w.repoGroundedness * clamp01(score.repoGroundedness) + w.driftPenalty * clamp01(score.driftPenalty) + w.toolUseQuality * clamp01(score.toolUseQuality) + w.patchQuality * clamp01(score.patchQuality) + w.testReality * clamp01(score.testReality) + w.finalGate * clamp01(score.finalGate) + w.reviewerBlockers * clamp01(score.reviewerBlockers) + w.costUsd * Math.max(0, finiteOrZero(score.costUsd)) + w.wallSeconds * Math.max(0, finiteOrZero(score.wallSeconds) / 60);
+}
+function clamp01(value) {
+  if (!Number.isFinite(value)) return 0;
+  return Math.max(0, Math.min(1, value));
+}
+function finiteOrZero(value) {
+  return Number.isFinite(value) ? value : 0;
+}
+
+export {
+  Mutex,
+  DEFAULT_RUN_SCORE_WEIGHTS,
+  aggregateRunScore,
+  clamp01
+};
+//# sourceMappingURL=chunk-MPHTT5HE.js.map

package/dist/chunk-MPHTT5HE.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/concurrency.ts","../src/run-score.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","export interface RunScore {\n  success: number\n  goalProgress: number\n  repoGroundedness: number\n  driftPenalty: number\n  toolUseQuality: number\n  patchQuality: number\n  testReality: number\n  finalGate: number\n  reviewerBlockers: number\n  costUsd: number\n  wallSeconds: number\n  notes?: string[]\n}\n\nexport interface RunScoreWeights {\n  success: number\n  goalProgress: number\n  repoGroundedness: number\n  driftPenalty: number\n  toolUseQuality: number\n  patchQuality: number\n  testReality: number\n  finalGate: number\n  reviewerBlockers: number\n  costUsd: number\n  wallSeconds: number\n}\n\nexport const DEFAULT_RUN_SCORE_WEIGHTS: RunScoreWeights = {\n  success: 4,\n  goalProgress: 2,\n  repoGroundedness: 1.5,\n  driftPenalty: -1.5,\n  toolUseQuality: 1,\n  patchQuality: 1.25,\n  testReality: 1.5,\n  finalGate: 3,\n  reviewerBlockers: -2,\n  costUsd: -0.2,\n  wallSeconds: -0.1,\n}\n\nexport function aggregateRunScore(score: RunScore, weights: Partial<RunScoreWeights> = {}): number {\n  const w = { ...DEFAULT_RUN_SCORE_WEIGHTS, ...weights }\n  return (\n    w.success * clamp01(score.success) +\n    w.goalProgress * clamp01(score.goalProgress) +\n    w.repoGroundedness * clamp01(score.repoGroundedness) +\n    w.driftPenalty * clamp01(score.driftPenalty) +\n    w.toolUseQuality * clamp01(score.toolUseQuality) +\n    w.patchQuality * clamp01(score.patchQuality) +\n    w.testReality * clamp01(score.testReality) +\n    w.finalGate * clamp01(score.finalGate) +\n    w.reviewerBlockers * clamp01(score.reviewerBlockers) +\n    w.costUsd * Math.max(0, finiteOrZero(score.costUsd)) +\n    w.wallSeconds * Math.max(0, finiteOrZero(score.wallSeconds) / 60)\n  )\n}\n\nexport function clamp01(value: number): number {\n  if (!Number.isFinite(value)) return 0\n  return Math.max(0, Math.min(1, value))\n}\n\nfunction finiteOrZero(value: number): number {\n  return Number.isFinite(value) ? value : 0\n}\n"],"mappings":";AAYO,IAAM,QAAN,MAAY;AAAA,EACT,SAAS;AAAA,EACA,UAA6B,CAAC;AAAA,EAE/C,MAAM,UAA+B;AACnC,QAAI,CAAC,KAAK,QAAQ;AAChB,WAAK,SAAS;AACd,aAAO,MAAM,KAAK,QAAQ;AAAA,IAC5B;AACA,WAAO,IAAI,QAAoB,CAAC,YAAY;AAC1C,WAAK,QAAQ,KAAK,MAAM;AACtB,gBAAQ,MAAM,KAAK,QAAQ,CAAC;AAAA,MAC9B,CAAC;AAAA,IACH,CAAC;AAAA,EACH;AAAA,EAEQ,UAAgB;AACtB,UAAM,OAAO,KAAK,QAAQ,MAAM;AAChC,QAAI,MAAM;AACR,WAAK;AAAA,IACP,OAAO;AACL,WAAK,SAAS;AAAA,IAChB;AAAA,EACF;AAAA,EAEA,MAAM,aAAgB,IAAsC;AAC1D,UAAM,UAAU,MAAM,KAAK,QAAQ;AACnC,QAAI;AACF,aAAO,MAAM,GAAG;AAAA,IAClB,UAAE;AACA,cAAQ;AAAA,IACV;AAAA,EACF;AAAA;AAAA,EAGA,IAAI,WAAoB;AACtB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA,EAGA,IAAI,UAAkB;AACpB,WAAO,KAAK,QAAQ;AAAA,EACtB;AACF;;;AC1BO,IAAM,4BAA6C;AAAA,EACxD,SAAS;AAAA,EACT,cAAc;AAAA,EACd,kBAAkB;AAAA,EAClB,cAAc;AAAA,EACd,gBAAgB;AAAA,EAChB,cAAc;AAAA,EACd,aAAa;AAAA,EACb,WAAW;AAAA,EACX,kBAAkB;AAAA,EAClB,SAAS;AAAA,EACT,aAAa;AACf;AAEO,SAAS,kBAAkB,OAAiB,UAAoC,CAAC,GAAW;AACjG,QAAM,IAAI,EAAE,GAAG,2BAA2B,GAAG,QAAQ;AACrD,SACE,EAAE,UAAU,QAAQ,MAAM,OAAO,IACjC,EAAE,eAAe,QAAQ,MAAM,YAAY,IAC3C,EAAE,mBAAmB,QAAQ,MAAM,gBAAgB,IACnD,EAAE,eAAe,QAAQ,MAAM,YAAY,IAC3C,EAAE,iBAAiB,QAAQ,MAAM,cAAc,IAC/C,EAAE,eAAe,QAAQ,MAAM,YAAY,IAC3C,EAAE,cAAc,QAAQ,MAAM,WAAW,IACzC,EAAE,YAAY,QAAQ,MAAM,SAAS,IACrC,EAAE,mBAAmB,QAAQ,MAAM,gBAAgB,IACnD,EAAE,UAAU,KAAK,IAAI,GAAG,aAAa,MAAM,OAAO,CAAC,IACnD,EAAE,cAAc,KAAK,IAAI,GAAG,aAAa,MAAM,WAAW,IAAI,EAAE;AAEpE;AAEO,SAAS,QAAQ,OAAuB;AAC7C,MAAI,CAAC,OAAO,SAAS,KAAK,EAAG,QAAO;AACpC,SAAO,KAAK,IAAI,GAAG,KAAK,IAAI,GAAG,KAAK,CAAC;AACvC;AAEA,SAAS,aAAa,OAAuB;AAC3C,SAAO,OAAO,SAAS,KAAK,IAAI,QAAQ;AAC1C;","names":[]}

package/dist/{chunk-WSBUZMBU.js → chunk-NBSS5NDZ.js}

@@ -2,10 +2,10 @@ import {
   Mutex,
   aggregateRunScore,
   clamp01
-} from "./chunk-QMXXSNC4.js";
+} from "./chunk-MPHTT5HE.js";
 import {
   computeFindingId
-} from "./chunk-5S5NJ63F.js";
+} from "./chunk-CIUOICJT.js";
 import {
   LlmClient,
   callLlmJson
@@ -820,4 +820,4 @@ export {
   runSemanticConceptJudge,
   createSemanticConceptJudge
 };
-//# sourceMappingURL=chunk-WSBUZMBU.js.map
+//# sourceMappingURL=chunk-NBSS5NDZ.js.map

package/dist/{code-agent-session-D-g04tcy.d.ts → code-agent-session-CdxteG0y.d.ts}

@@ -1,4 +1,4 @@
-import { b as RunSplitTag, c as RunCostProvenance, R as RunRecord } from './run-record-B7RTi_ix.js';
+import { a as RunSplitTag, b as RunCostProvenance, R as RunRecord } from './run-record-CZmcpWPo.js';
 
 type CodeAgentSessionSource = 'codex' | 'claude-code' | 'opencode' | 'kimi-code' | 'pi';
 interface ParsedCodeAgentJsonl {

package/dist/contract/index.d.ts

@@ -1,21 +1,21 @@
-import { S as Scenario, M as MutableSurface, D as DispatchContext, b as JudgeConfig, f as SurfaceProposer, g as Gate, L as LabeledScenarioStore, C as CampaignResult, h as GateDecision } from '../types-C5gJrOVT.js';
-export { i as CampaignAggregates, j as CampaignArtifactWriter, k as CampaignCellResult, l as CampaignCostMeter, d as CampaignTraceWriter, m as CodeSurface, c as Dispatch, n as GateContext, G as GateResult, o as GenerationCandidate, e as GenerationRecord, a as JudgeDimension, J as JudgeScore, p as Mutator, O as OptimizationProposer, q as OptimizerConfig, r as SessionScript } from '../types-C5gJrOVT.js';
-import { L as LoopProvenanceRecord, P as PowerPreflight, R as RunEvalOptions } from '../provenance-Bibyg1U9.js';
-export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, O as ObjectiveSource, c as ParetoSignificanceGateOptions, d as PromotionObjective, e as PromotionPolicy, f as buildEvidenceVector, g as composeGate, h as defaultProductionGate, i as evolutionaryProposer, j as heldOutGate, p as paretoPolicy, k as paretoSignificanceGate, r as runEval } from '../provenance-Bibyg1U9.js';
-import { R as RunOptimizationOptions, a as RunImprovementLoopResult } from '../gepa-dne9JDPL.js';
-export { G as GepaProposerOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions, g as gepaProposer, r as runCampaign, d as runImprovementLoop } from '../gepa-dne9JDPL.js';
+import { S as Scenario, M as MutableSurface, D as DispatchContext, b as JudgeConfig, f as SurfaceProposer, g as Gate, L as LabeledScenarioStore, C as CampaignResult, h as GateDecision } from '../types-Ca_63YSD.js';
+export { i as CampaignAggregates, j as CampaignArtifactWriter, k as CampaignCellResult, l as CampaignCostMeter, d as CampaignTraceWriter, m as CodeSurface, c as Dispatch, n as GateContext, G as GateResult, o as GenerationCandidate, e as GenerationRecord, a as JudgeDimension, J as JudgeScore, p as Mutator, O as OptimizationProposer, q as OptimizerConfig, r as SessionScript } from '../types-Ca_63YSD.js';
+import { L as LoopProvenanceRecord, P as PowerPreflight, R as RunEvalOptions } from '../provenance-BbVagC68.js';
+export { A as AxisEvidence, a as AxisVerdict, B as BuildEvidenceVectorOptions, D as DefaultProductionGateOptions, E as EvidenceVector, b as EvolutionaryProposerOptions, H as HeldOutGateOptions, O as ObjectiveSource, c as ParetoSignificanceGateOptions, d as PromotionObjective, e as PromotionPolicy, f as buildEvidenceVector, g as composeGate, h as defaultProductionGate, i as evolutionaryProposer, j as heldOutGate, p as paretoPolicy, k as paretoSignificanceGate, r as runEval } from '../provenance-BbVagC68.js';
+import { R as RunOptimizationOptions, a as RunImprovementLoopResult } from '../gepa-CQelRtuC.js';
+export { G as GepaProposerOptions, b as RunCampaignOptions, c as RunImprovementLoopOptions, g as gepaProposer, r as runCampaign, d as runImprovementLoop } from '../gepa-CQelRtuC.js';
 import { C as CampaignStorage } from '../storage-Dw_f7WMt.js';
 export { f as fsCampaignStorage, i as inMemoryCampaignStorage } from '../storage-Dw_f7WMt.js';
 export { D as DeploymentOutcome, F as FileSystemOutcomeStore, a as FileSystemOutcomeStoreOptions, I as InMemoryOutcomeStore, b as OutcomeStore } from '../outcome-store-rnXLEqSn.js';
 import { HostedTenant, EvalRunCellScore, EvalRunGenerationSnapshot, EvalRunEvent, TraceSpanEvent } from '../hosted/index.js';
-import { R as RunRecord, b as RunSplitTag } from '../run-record-B7RTi_ix.js';
-import { I as InsightReport } from '../insight-report-IwwvqZZv.js';
-export { C as CostProvenanceSummary, F as FailureClusterInsight, a as InterRaterInsight, J as JudgeInsight, L as LiftInsight, O as OutcomeCorrelationInsight, R as Recommendation, b as ReleaseSummary, S as ScalarDistribution } from '../insight-report-IwwvqZZv.js';
-export { D as DefaultAnalystRegistryOptions, c as buildDefaultAnalystRegistry } from '../default-registry-DltpYR5u.js';
-export { A as AnalystFinding } from '../kind-factory-DcNg13sZ.js';
-import { A as AnalyzeRunsOptions } from '../analyze-runs-BYHg6Irm.js';
-export { a as analyzeRuns } from '../analyze-runs-BYHg6Irm.js';
-export { C as CodeAgentSessionDiagnostic, a as CodeAgentSessionIntakeOptions, b as CodeAgentSessionIntakeResult, c as CodeAgentSessionMetrics, d as CodeAgentSessionSource, P as ParsedCodeAgentJsonl, f as fromClaudeCodeSession, e as fromCodexSession, g as fromKimiCodeSession, h as fromOpenCodeSession, i as fromPiSession, j as fromPigraphSession, p as parseCodeAgentJsonl } from '../code-agent-session-D-g04tcy.js';
+import { R as RunRecord, a as RunSplitTag } from '../run-record-CZmcpWPo.js';
+import { I as InsightReport } from '../insight-report-oMVxDTxl.js';
+export { C as CostProvenanceSummary, F as FailureClusterInsight, a as InterRaterInsight, J as JudgeInsight, L as LiftInsight, O as OutcomeCorrelationInsight, R as Recommendation, b as ReleaseSummary, S as ScalarDistribution } from '../insight-report-oMVxDTxl.js';
+export { D as DefaultAnalystRegistryOptions, c as buildDefaultAnalystRegistry } from '../default-registry-DDfv22MQ.js';
+export { d as AnalystFinding } from '../policy-edit-Clb2v6Oa.js';
+import { A as AnalyzeRunsOptions } from '../analyze-runs-0rz_m29H.js';
+export { a as analyzeRuns } from '../analyze-runs-0rz_m29H.js';
+export { C as CodeAgentSessionDiagnostic, a as CodeAgentSessionIntakeOptions, b as CodeAgentSessionIntakeResult, c as CodeAgentSessionMetrics, d as CodeAgentSessionSource, P as ParsedCodeAgentJsonl, f as fromClaudeCodeSession, e as fromCodexSession, g as fromKimiCodeSession, h as fromOpenCodeSession, i as fromPiSession, j as fromPigraphSession, p as parseCodeAgentJsonl } from '../code-agent-session-CdxteG0y.js';
 import '../statistics-oUbOJe-S.js';
 import '../judge-calibration-7C-IDmKr.js';
 import '../types-C7DGg5ex.js';
@@ -24,13 +24,12 @@ import '../dataset-NENEzRgk.js';
 import '../errors-oeQrLqXC.js';
 import '../store-BsVi7ncX.js';
 import '../schema-SGWcK9wa.js';
-import '../llm-client-DyqEH4jH.js';
-import '../raw-provider-sink-C46HDghv.js';
 import '@tangle-network/agent-interface';
-import '../summary-report-BJ5aNwZ1.js';
+import '../summary-report-DTNgQycC.js';
 import '../failure-cluster-C48PiReX.js';
 import '@ax-llm/ax';
-import '../store-C1YxJDEK.js';
+import '../kind-factory-DWOvXjR_.js';
+import '../store-9cAScOcb.js';
 import 'zod';
 
 /**
@@ -75,7 +74,8 @@ interface SelfImproveBudget {
     holdoutScenarios?: Scenario[];
     /** Per-scenario replicates per cell — raises bootstrap-CI tightness. Default 1. */
     reps?: number;
-    /** Top-scoring candidates carried into the next generation. Default 2. */
+    /** @deprecated Must be 1 when supplied. The loop promotes only a candidate
+     *  that replaces its single global incumbent. */
     promoteTopK?: number;
 }
 interface SelfImproveLlm {

package/dist/contract/index.js

@@ -15,6 +15,7 @@ import {
 } from "../chunk-J6P6PK2R.js";
 import {
   buildEvidenceVector,
+  campaignMeanComposite,
   composeGate,
   defaultProductionGate,
   emitLoopProvenance,
@@ -28,7 +29,7 @@ import {
   runImprovementLoop,
   surfaceContentHash,
   surfaceHash
-} from "../chunk-ADYLPOSX.js";
+} from "../chunk-3274WNK7.js";
 import "../chunk-VI2UW6B6.js";
 import {
   fsCampaignStorage,
@@ -37,8 +38,8 @@ import {
 } from "../chunk-FAOEFFRT.js";
 import {
   buildDefaultAnalystRegistry
-} from "../chunk-I6LVHOV3.js";
-import "../chunk-5S5NJ63F.js";
+} from "../chunk-7GKEAIAD.js";
+import "../chunk-CIUOICJT.js";
 import {
   FileSystemOutcomeStore,
   InMemoryOutcomeStore
@@ -223,6 +224,7 @@ async function selfImprove(opts) {
     winnerLabel: result.winnerLabel,
     winnerRationale: result.winnerRationale,
     diff: result.promotedDiff,
+    baselineSearchComposite: campaignMeanComposite(result.baselineCampaign),
     generations: result.generations.map((g) => ({
       generationIndex: g.record.generationIndex,
       candidates: g.record.candidates,