agent-afk 5.256.0 → 5.257.0

package/dist/whatif/types.d.ts

@@ -0,0 +1,260 @@
+export type Change = {
+    kind: 'append';
+    target: 'user-afk-md' | 'project-afk-md';
+    text: string;
+} | {
+    kind: 'file';
+    path: string;
+    content: string;
+} | {
+    kind: 'hot';
+    content: string;
+} | {
+    kind: 'memory-add';
+    content: string;
+    category: 'preference' | 'convention' | 'decision' | 'learning';
+} | {
+    kind: 'memory-remove';
+    id: number;
+} | {
+    kind: 'disable-skill';
+    name: string;
+} | {
+    kind: 'disable-plugin';
+    name: string;
+} | {
+    kind: 'model';
+    model: string;
+} | {
+    kind: 'effort';
+    effort: string;
+} | {
+    kind: 'env';
+    key: string;
+    value: string;
+};
+export type ChangeKind = Change['kind'];
+export interface ChangeSpec {
+    title: string;
+    changes: Change[];
+}
+export interface LaunchSettings {
+    model?: string;
+    effort?: string;
+    env: Record<string, string>;
+    unset?: string[];
+}
+export interface Environment {
+    label: 'baseline' | 'candidate';
+    home: string;
+    cwd: string;
+    launch: LaunchSettings;
+}
+export interface OperatorContext {
+    realHome: string;
+    realCwd: string;
+}
+export interface ChangeOperator<K extends ChangeKind = ChangeKind> {
+    kind: K;
+    touchesProject(change: Extract<Change, {
+        kind: K;
+    }>): boolean;
+    homePathsToCopy(change: Extract<Change, {
+        kind: K;
+    }>): string[];
+    apply(change: Extract<Change, {
+        kind: K;
+    }>, env: Environment, ctx: OperatorContext): Promise<void>;
+    describe(change: Extract<Change, {
+        kind: K;
+    }>): string;
+}
+export interface RequestSnapshot {
+    model: string;
+    system: string;
+    tools: {
+        name: string;
+        description: string;
+    }[];
+    firstUserMessage: string;
+}
+export interface StructuralImpact {
+    baseline: RequestSnapshot;
+    candidate: RequestSnapshot;
+    systemDiff: string;
+    toolsAdded: string[];
+    toolsRemoved: string[];
+    toolsChanged: string[];
+    userMessageDiff: string;
+    tokens: {
+        baseline: number;
+        candidate: number;
+    };
+    perTurnCostDeltaUsd?: number;
+    modelChanged: boolean;
+}
+export type PredictionDirection = 'added' | 'removed' | 'strengthened' | 'weakened';
+export type Confidence = 'high' | 'medium' | 'low';
+export interface Prediction {
+    id: string;
+    behavior: string;
+    direction: PredictionDirection;
+    confidence: Confidence;
+    reason: string;
+    testQuestion: string;
+    probes: string[];
+}
+export interface Episode {
+    id: string;
+    source: 'real' | 'synthetic' | 'suite';
+    prompt: string;
+    targets?: string;
+}
+export interface ToolRequest {
+    tool: string;
+    input: unknown;
+    verdict: 'executed' | 'recorded';
+}
+export interface EpisodeTrace {
+    episodeId: string;
+    env: 'baseline' | 'candidate';
+    sample: number;
+    text: string;
+    tools: ToolRequest[];
+    costUsd: number;
+    inputTokens: number;
+    outputTokens: number;
+    durationMs: number;
+    error?: string;
+}
+export interface RunnerOptions {
+    timeoutMs: number;
+    maxTurns: number;
+    signal?: AbortSignal;
+}
+export interface AgentRunner {
+    name: string;
+    run(env: Environment, episode: Episode, sample: number, opts: RunnerOptions): Promise<EpisodeTrace>;
+    snapshot(env: Environment, probePrompt: string, opts: RunnerOptions): Promise<RequestSnapshot>;
+}
+export interface EpisodeFeatures {
+    firstAction: 'answer' | 'read' | 'side-effect' | 'ask' | 'delegate' | 'none';
+    askedBeforeActing: boolean;
+    delegated: boolean;
+    usedTools: string[];
+    usedSkills: string[];
+    searchedMemory: boolean;
+    toolCalls: number;
+    responseChars: number;
+    errored: boolean;
+}
+export interface JudgeQuestion {
+    id: string;
+    question: string;
+}
+export interface JudgeInput {
+    prompt: string;
+    output: string;
+    questions: JudgeQuestion[];
+}
+export type JudgeResult = Record<string, number>;
+export interface Judge {
+    name: 'jev' | 'claude';
+    external: boolean;
+    grade(input: JudgeInput, signal?: AbortSignal): Promise<JudgeResult>;
+    close?(): Promise<void>;
+}
+export interface RateComparison {
+    baseline: number;
+    candidate: number;
+    delta: number;
+    ci: [number, number];
+    n: {
+        baseline: number;
+        candidate: number;
+    };
+}
+export type Verdict = 'confirmed' | 'refuted' | 'unclear';
+export interface VerifiedPrediction {
+    prediction: Prediction;
+    rates: RateComparison;
+    verdict: Verdict;
+}
+export interface DiscoveredDifference {
+    id: string;
+    description: string;
+    question: string;
+    rates: RateComparison;
+}
+export interface FeatureDelta {
+    label: string;
+    rates: RateComparison;
+}
+export interface VerifyResult {
+    predictions: VerifiedPrediction[];
+    discovered: DiscoveredDifference[];
+    features: FeatureDelta[];
+    episodes: number;
+    samples: number;
+    judge: {
+        name: 'jev' | 'claude';
+        external: boolean;
+        crossCheckAgreement?: number;
+    };
+    predictionAccuracy?: number;
+    truncatedByBudget: boolean;
+    failedEpisodes: number;
+    judgeFailures?: number;
+}
+export interface WhatifReport {
+    spec: ChangeSpec;
+    structural: StructuralImpact;
+    predictions: Prediction[];
+    verify?: VerifyResult;
+    costUsd: number;
+    runDir: string;
+    limits: string[];
+    headline: string;
+}
+export type WhatifStage = 'compile' | 'sandbox' | 'snapshot' | 'predict' | 'episodes' | 'run' | 'judge' | 'discover' | 'report';
+export interface WhatifProgress {
+    stage: WhatifStage;
+    message: string;
+    done?: number;
+    total?: number;
+}
+export type CompleteFn = (req: {
+    system: string;
+    user: string;
+    maxTokens: number;
+    model: string;
+    signal?: AbortSignal;
+}) => Promise<{
+    text: string;
+    costUsd: number;
+}>;
+export interface WhatifOptions {
+    spec: ChangeSpec;
+    realHome: string;
+    realCwd: string;
+    agentModel: string;
+    analystModel: string;
+    verify: boolean;
+    turns: number;
+    samples: number;
+    maxUsd: number;
+    judge: 'auto' | 'jev' | 'claude';
+    concurrency: number;
+    maxTurns: number;
+    episodeTimeoutMs: number;
+    keepSandboxes: boolean;
+}
+export interface WhatifDeps {
+    runner: AgentRunner;
+    complete: CompleteFn;
+    makeJudge(choice: WhatifOptions['judge']): Promise<Judge>;
+    makeCrossCheckJudge(): Promise<Judge | undefined>;
+    onProgress?: (p: WhatifProgress) => void;
+    signal?: AbortSignal;
+    now?: () => Date;
+}

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-afk",
-  "version": "5.256.0",
+  "version": "5.257.0",
   "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",