@tangle-network/agent-eval 0.95.0 → 0.96.0

package/dist/{types.d.ts → types-C7DGg5ex.d.ts}

@@ -1,4 +1,6 @@
-export interface Scenario {
+import { TCloud } from '@tangle-network/tcloud';
+
+interface Scenario {
     id: string;
     persona: string;
     label: string;
@@ -8,37 +10,37 @@ export interface Scenario {
     artifactChecks: ArtifactCheck[];
     systemPromptAppend?: string;
 }
-export interface Turn {
+interface Turn {
     user: string;
     expectedBehaviors: string[];
     adversarial?: boolean;
     feedbackType?: 'correction' | 'rejection' | 'vague' | 'contradictory' | 'escalation';
 }
-export interface ArtifactCheck {
+interface ArtifactCheck {
     type: 'vault_file_exists' | 'vault_file_contains' | 'block_extracted' | 'code_valid' | 'generation_produced' | 'tool_created' | string;
     target: string;
     contains?: string;
     minCount?: number;
     description: string;
 }
-export interface JudgeConfig {
+interface JudgeConfig {
     model: string;
     temperature: number;
     rubric: JudgeRubric;
 }
-export interface JudgeRubric {
+interface JudgeRubric {
     name: string;
     description: string;
     dimensions: RubricDimension[];
 }
-export interface RubricDimension {
+interface RubricDimension {
     name: string;
     description: string;
     anchor_low: string;
     anchor_high: string;
     weight: number;
 }
-export interface ScenarioResult {
+interface ScenarioResult {
     scenarioId: string;
     persona: string;
     turns: TurnResult[];
@@ -49,7 +51,7 @@ export interface ScenarioResult {
     totalDurationMs: number;
     artifacts: CollectedArtifacts;
 }
-export interface TurnResult {
+interface TurnResult {
     turnIndex: number;
     userMessage: string;
     agentResponse: string;
@@ -61,19 +63,19 @@ export interface TurnResult {
     containsCode: boolean;
     containsToolCall: boolean;
 }
-export interface ArtifactResult {
+interface ArtifactResult {
     check: ArtifactCheck;
     passed: boolean;
     detail?: string;
 }
-export interface JudgeScore {
+interface JudgeScore {
     judgeName: string;
     dimension: string;
     score: number;
     reasoning: string;
     evidence?: string;
 }
-export interface CollectedArtifacts {
+interface CollectedArtifacts {
     vaultFiles: {
         path: string;
         content: string;
@@ -88,7 +90,7 @@ export interface CollectedArtifacts {
     }[];
     toolCalls: string[];
 }
-export interface BenchmarkReport {
+interface BenchmarkReport {
     timestamp: string;
     generation: number;
     promptVersion: string;
@@ -117,7 +119,7 @@ export interface BenchmarkReport {
         }[];
     };
 }
-export interface RouteMap {
+interface RouteMap {
     signup?: string;
     login?: string;
     workspaces?: string;
@@ -130,13 +132,13 @@ export interface RouteMap {
     generations?: string;
     [key: string]: string | undefined;
 }
-export interface ProductClientConfig {
+interface ProductClientConfig {
     baseUrl: string;
     routes: RouteMap;
     /** Per-request timeout in ms before the request is aborted. Default 30s. */
     timeoutMs?: number;
 }
-export interface ScenarioFile {
+interface ScenarioFile {
     id: string;
     category: string;
     persona: string;
@@ -153,12 +155,12 @@ export interface ScenarioFile {
     turns: Turn[];
     artifactChecks: ArtifactCheck[];
 }
-export interface CompletionCriterion {
+interface CompletionCriterion {
     name: string;
     check: (state: DriverState) => boolean;
     progress?: (state: DriverState) => number;
 }
-export interface FeedbackPattern {
+interface FeedbackPattern {
     trigger: string;
     response: string;
 }
@@ -170,8 +172,8 @@ export interface FeedbackPattern {
  *   relentless  — senior partner reviewing for a client who will litigate;
  *                 interrogates every claim, accepts nothing undefended.
  */
-export type PersonaRigor = 'cooperative' | 'demanding' | 'relentless';
-export interface PersonaConfig {
+type PersonaRigor = 'cooperative' | 'demanding' | 'relentless';
+interface PersonaConfig {
     id: string;
     role: string;
     goal: string;
@@ -202,7 +204,7 @@ export interface PersonaConfig {
      */
     curveballs?: string[];
 }
-export interface DriverState {
+interface DriverState {
     tasks: number;
     events: number;
     proposals: {
@@ -214,7 +216,7 @@ export interface DriverState {
     codeBlocks: number;
     generations: number;
 }
-export interface TurnMetrics {
+interface TurnMetrics {
     turn: number;
     timestamp: string;
     tasks: number;
@@ -236,7 +238,7 @@ export interface TurnMetrics {
     totalCostUsd: number;
     completionPercent: number;
 }
-export interface DriverResult {
+interface DriverResult {
     personaId: string;
     /** True when the simulated user professionally signed off (driver said DONE). */
     completed: boolean;
@@ -255,7 +257,7 @@ export interface DriverResult {
     totalCostUsd: number;
     finalQualityScore: number | null;
 }
-export interface BenchmarkRunnerConfig {
+interface BenchmarkRunnerConfig {
     scenarios: Scenario[];
     judges: JudgeFn[];
     systemPrompt: string;
@@ -265,32 +267,32 @@ export interface BenchmarkRunnerConfig {
     generation?: number;
     promptVersion?: string;
 }
-export interface JudgeInput {
+interface JudgeInput {
     scenario: Scenario;
     turns: TurnResult[];
     artifacts: CollectedArtifacts;
 }
-export type JudgeFn = (tc: TCloud, input: JudgeInput) => Promise<JudgeScore[]>;
-import type { TCloud } from '@tangle-network/tcloud';
-export type { TCloud };
-export interface TestResult {
+type JudgeFn = (tc: TCloud, input: JudgeInput) => Promise<JudgeScore[]>;
+
+interface TestResult {
     name: string;
     passed: boolean;
     duration: number;
     detail?: string;
     checks: CheckResult[];
 }
-export interface CheckResult {
+interface CheckResult {
     name: string;
     passed: boolean;
     expected: string;
     actual: string;
 }
-export interface EvalResult {
+interface EvalResult {
     scenario: string;
     status: 'pass' | 'fail' | 'skip';
     duration: number;
     detail?: string;
     artifact?: string;
 }
-//# sourceMappingURL=types.d.ts.map
+
+export type { ArtifactCheck as A, BenchmarkRunnerConfig as B, CheckResult as C, DriverResult as D, EvalResult as E, FeedbackPattern as F, JudgeScore as J, ProductClientConfig as P, RouteMap as R, Scenario as S, TestResult as T, JudgeInput as a, JudgeFn as b, BenchmarkReport as c, PersonaConfig as d, DriverState as e, CollectedArtifacts as f, ScenarioResult as g, TurnMetrics as h, ScenarioFile as i, CompletionCriterion as j, ArtifactResult as k, JudgeConfig as l, JudgeRubric as m, PersonaRigor as n, RubricDimension as o, Turn as p, TurnResult as q };

package/dist/{verdict.d.ts → verdict-C9MlYujm.d.ts}

@@ -21,7 +21,7 @@
  * Need structured extras? Extend DefaultVerdict with typed fields — never
  * serialize extras into `notes`.
  */
-export interface DefaultVerdict {
+interface DefaultVerdict {
     /** Whether the output meets the validator's pass criteria. */
     valid: boolean;
     /** Aggregate score in [0, 1]. Drivers use this for winner selection. */
@@ -31,4 +31,5 @@ export interface DefaultVerdict {
     /** Human-readable rationale; surfaces in trace + final-result `winner.verdict`. */
     notes?: string;
 }
-//# sourceMappingURL=verdict.d.ts.map
+
+export type { DefaultVerdict as D };