agent-afk 5.268.0 → 5.269.0

package/dist/whatif/probe-signflip.d.ts

@@ -0,0 +1,17 @@
+export declare const ZERO_TOLERANCE = 1e-9;
+export interface ProbeSignFlipResult {
+    nPaired: number;
+    nUnpaired: number;
+    nNonzero: number;
+    meanDelta: number;
+    probeDiffs: number[];
+    p: number | null;
+    minAchievableP: number | null;
+    underpoweredForSig: boolean;
+    method: 'exact' | 'montecarlo';
+}
+export interface ArmSamples {
+    baseline: number[];
+    candidate: number[];
+}
+export declare function computeProbeSignFlip(pairedEpisodeData: Map<string, ArmSamples>, episodeOrder: string[]): ProbeSignFlipResult;

package/dist/whatif/report.predictions.d.ts

@@ -0,0 +1,3 @@
+import type { Prediction, VerifiedPrediction } from './types.js';
+export declare function scoredOn(vp: VerifiedPrediction): string;
+export declare function renderVerifiedPredictionTable(predictions: Prediction[], verified: VerifiedPrediction[]): string[];

package/dist/whatif/report.signflip.d.ts

@@ -0,0 +1,4 @@
+import type { ArmImbalance } from './run.failures.js';
+import type { VerifiedPrediction } from './types.js';
+export declare function fmtP(p: number): string;
+export declare function renderProbeSignFlipSection(verified: VerifiedPrediction[], armImbalance?: ArmImbalance): string[];

package/dist/whatif/run.failures.d.ts

@@ -0,0 +1,22 @@
+import type { EpisodeTrace } from './types.js';
+export interface FailedEpisodeRecord {
+    episodeId: string;
+    arm: 'baseline' | 'candidate';
+    sample: number;
+    errorClass: 'timeout' | 'error';
+    errorMessage: string;
+    probe?: string;
+    durationMs: number;
+}
+export interface ArmImbalance {
+    baselineFailRate: number;
+    candidateFailRate: number;
+    rateDiff: number;
+    allInOneArm: boolean;
+    concentrationArm?: 'baseline' | 'candidate';
+    summary: string;
+}
+export declare const IMBALANCE_RATE_THRESHOLD = 0.2;
+export declare const IMBALANCE_MIN_FAILURES = 2;
+export declare function buildFailedEpisodeRecords(traces: EpisodeTrace[], episodeTargets: Map<string, string>): FailedEpisodeRecord[];
+export declare function detectArmImbalance(records: FailedEpisodeRecord[], baselineTotal: number, candidateTotal: number): ArmImbalance | undefined;

package/dist/whatif/run.persist.d.ts

@@ -1,2 +1,3 @@
 import type { EpisodeTrace, WhatifReport } from './types.js';
-export declare function persistRun(runDir: string, report: WhatifReport, traces: EpisodeTrace[]): Promise<void>;
+import type { JudgeResults } from './run.verify.scoring.js';
+export declare function persistRun(runDir: string, report: WhatifReport, traces: EpisodeTrace[], judgeResults?: JudgeResults): Promise<void>;

package/dist/whatif/run.persist.grades.d.ts

@@ -0,0 +1,10 @@
+import type { EpisodeTrace } from './types.js';
+import type { JudgeResults } from './run.verify.scoring.js';
+export interface GradeEntry {
+    episodeId: string;
+    env: 'baseline' | 'candidate';
+    sample: number;
+    predictionId: string;
+    pYes: number;
+}
+export declare function persistGrades(runDir: string, judgeResults: JudgeResults, traces: EpisodeTrace[], predictionIds: string[]): Promise<void>;

package/dist/whatif/run.report.d.ts

@@ -2,6 +2,7 @@ import type { ChangeSpec, EpisodeTrace, Judge, Prediction, VerifyResult, WhatifR
 import type { DroppedProbe } from './probe-grounding.js';
 import type { CorpusExclusions } from './episodes.js';
 import type { StructuralImpact } from './types.js';
+import type { JudgeResults } from './run.verify.scoring.js';
 export interface VerifiedReportArgs {
     spec: ChangeSpec;
     structural: StructuralImpact;
@@ -14,5 +15,6 @@ export interface VerifiedReportArgs {
     runDir: string;
     resolvedJudge: Judge;
     autoKeepContextHooks: boolean;
+    judgeResults: JudgeResults;
 }
 export declare function buildAndPersistVerifiedReport(args: VerifiedReportArgs): Promise<WhatifReport>;

package/dist/whatif/run.verify.d.ts

@@ -1,3 +1,4 @@
+import { type JudgeResults } from './run.verify.scoring.js';
 import type { AgentRunner, CompleteFn, Episode, EpisodeTrace, Environment, Judge, Prediction, StructuralImpact, VerifyResult, WhatifProgress } from './types.js';
 export interface VerifyRunInput {
     episodes: Episode[];
@@ -26,5 +27,6 @@ export interface VerifyRunOutput {
     verifyResult: VerifyResult;
     allTraces: EpisodeTrace[];
     analystCostUsd: number;
+    judgeResults: JudgeResults;
 }
 export declare function verifyRun(input: VerifyRunInput): Promise<VerifyRunOutput>;

package/dist/whatif/types.d.ts

@@ -196,6 +196,7 @@ export interface VerifiedPrediction {
     crossCheckAgreement?: number;
     crossCheckTooFew?: boolean;
     verdictReason?: 'judges disagree';
+    probeSignFlip?: import('./probe-signflip.js').ProbeSignFlipResult;
 }
 export interface DiscoveredDifference {
     id: string;
@@ -222,6 +223,8 @@ export interface VerifyResult {
     truncatedByBudget: boolean;
     failedEpisodes: number;
     judgeFailures?: number;
+    failedEpisodeRecords?: import('./run.failures.js').FailedEpisodeRecord[];
+    armImbalance?: import('./run.failures.js').ArmImbalance;
 }
 export interface WhatifReport {
     spec: ChangeSpec;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-afk",
-  "version": "5.268.0",
+  "version": "5.269.0",
   "description": "Open-source coding-agent harness you can actually change — own the loop (prompts, gates, routing, skills, terminal states), use any model, run long tasks while you're away.",
   "main": "dist/index.mjs",
   "types": "dist/index.d.ts",