agent-afk 5.266.2 → 5.267.0

package/dist/whatif/args.d.ts

@@ -21,6 +21,7 @@ export interface ParsedWhatifArgs {
     options: WhatifFlagOptions;
     yes: boolean;
     json: boolean;
+    force: boolean;
 }
 export type ParseResult = ParsedWhatifArgs | string;
 export declare function parseWhatifArgs(argv: string[]): ParseResult;

package/dist/whatif/args.usage.d.ts

@@ -1 +1 @@
-export declare const WHATIF_USAGE = "afk whatif \u2014 predict the behavioural impact of a proposed change.\n\nUSAGE\n  afk whatif [change text...]  [flags]\n  /whatif [change text...]  [flags]\n\nEXAMPLES\n  # Plain English \u2014 compiled to a ChangeSpec, confirmed before running\n  afk whatif \"turn off auto-routing\"\n  afk whatif \"always ask before using tools\"\n\n  # Explicit flags (can be combined and repeated)\n  afk whatif --append \"Always ask a clarifying question before running tools.\"\n  afk whatif --model claude-haiku-4-5 --verify\n  afk whatif --memory-add \"prefers pnpm test:file\" --memory-category preference\n\n  # Load a prepared ChangeSpec file\n  afk whatif --spec ./my-change.json\n\n  # Quick single-turn estimate\n  afk whatif \"disable diagnose skill\" --quick\n\n  # Full verification with custom judge and budget\n  afk whatif --append \"Never use bash\" --verify --judge claude --max-usd 3\n\nCHANGE FLAGS (accumulate in order)\n  --append <text>              Append text to your AFK.md (user scope)\n  --append-project <text>      Append text to project AFK.md\n  --file <path>=<localfile>    Set a file (path: home:<rel> or project:<rel>)\n  --hot <localfile>            Replace HOT.md with a local file's content\n  --memory-add <text>          Add a memory fact\n  --memory-category <cat>      Category for the next --memory-add\n                               (preference|convention|decision|learning; default: preference)\n  --memory-remove <id>         Remove a memory fact by numeric id\n  --disable-skill <name>       Disable a skill by name\n  --disable-plugin <name>      Disable a plugin by name\n  --model <id>                 Test a different candidate model\n  --effort <level>             Test a different effort level\n  --env KEY=VALUE              Set an env var in the candidate sandbox\n  --spec <file.json>           Load a full ChangeSpec from a JSON file\n\nRUN OPTIONS\n  --agent-model <id>           Model the agent under test uses (default: current)\n  --analyst-model <id>         Model for compile/predict/judge (default: sonnet)\n  --verify                     Run episodes and verify predictions empirically\n  --quick                      Single-turn episodes (sets --max-turns 1)\n  --turns <n>                  Real turns to replay (default: 12)\n  --samples <n>                Samples per episode per environment (default: 3)\n  --max-usd <n>                Budget cap in USD (default: 5)\n  --judge auto|jev|claude      Judge to use (default: auto)\n  --concurrency <n>            Parallel episodes (default: 4)\n  --max-turns <n>              Max turns per episode (default: 3)\n  --timeout <sec>              Episode timeout in seconds (default: 180)\n  --keep-sandboxes             Keep sandbox directories after run\n  --yes                        Skip confirmation of compiled spec\n  --json                       Print results as JSON to stdout\n\nENV VARS (episode behaviour)\n  AFK_WHATIF_KEEP_CONTEXT_HOOKS=1\n    Keep SessionStart and UserPromptSubmit hooks enabled inside episodes.\n    By default those hooks are disabled so both arms see byte-identical\n    first user messages. Set this when your change specifically tests\n    context-injecting hook behaviour. The engine sets it automatically\n    when the change spec targets hooks or plugins (disable-plugin, or a\n    file change to afk.config.json or a hooks.json manifest).\n";
+export declare const WHATIF_USAGE = "afk whatif \u2014 predict the behavioural impact of a proposed change.\n\nUSAGE\n  afk whatif [change text...]  [flags]\n  /whatif [change text...]  [flags]\n\nEXAMPLES\n  # Plain English \u2014 compiled to a ChangeSpec, confirmed before running\n  afk whatif \"turn off auto-routing\"\n  afk whatif \"always ask before using tools\"\n\n  # Explicit flags (can be combined and repeated)\n  afk whatif --append \"Always ask a clarifying question before running tools.\"\n  afk whatif --model claude-haiku-4-5 --verify\n  afk whatif --memory-add \"prefers pnpm test:file\" --memory-category preference\n\n  # Load a prepared ChangeSpec file\n  afk whatif --spec ./my-change.json\n\n  # Quick single-turn estimate\n  afk whatif \"disable diagnose skill\" --quick\n\n  # Full verification with custom judge and budget\n  afk whatif --append \"Never use bash\" --verify --judge claude --max-usd 3\n\nCHANGE FLAGS (accumulate in order)\n  --append <text>              Append text to your AFK.md (user scope)\n  --append-project <text>      Append text to project AFK.md\n  --file <path>=<localfile>    Set a file (path: home:<rel> or project:<rel>)\n  --hot <localfile>            Replace HOT.md with a local file's content\n  --memory-add <text>          Add a memory fact\n  --memory-category <cat>      Category for the next --memory-add\n                               (preference|convention|decision|learning; default: preference)\n  --memory-remove <id>         Remove a memory fact by numeric id\n  --disable-skill <name>       Disable a skill by name\n  --disable-plugin <name>      Disable a plugin by name\n  --model <id>                 Test a different candidate model\n  --effort <level>             Test a different effort level\n  --env KEY=VALUE              Set an env var in the candidate sandbox\n  --spec <file.json>           Load a full ChangeSpec from a JSON file\n\nRUN OPTIONS\n  --agent-model <id>           Model the agent under test uses (default: current)\n  --analyst-model <id>         Model for compile/predict/judge (default: sonnet)\n  --verify                     Run episodes and verify predictions empirically\n  --quick                      Single-turn episodes (sets --max-turns 1)\n  --turns <n>                  Real turns to replay (default: 12)\n  --samples <n>                Samples per episode per environment (default: 3)\n  --max-usd <n>                Budget cap in USD (default: 5)\n  --judge auto|jev|claude      Judge to use (default: auto)\n  --concurrency <n>            Parallel episodes (default: 4)\n  --max-turns <n>              Max turns per episode (default: 3)\n  --timeout <sec>              Episode timeout in seconds (default: 180)\n  --keep-sandboxes             Keep sandbox directories after run\n  --yes                        Skip confirmation of compiled spec\n  --force                      Bypass the MDE underpowered gate (--verify only)\n  --json                       Print results as JSON to stdout\n\nENV VARS (episode behaviour)\n  AFK_WHATIF_KEEP_CONTEXT_HOOKS=1\n    Keep SessionStart and UserPromptSubmit hooks enabled inside episodes.\n    By default those hooks are disabled so both arms see byte-identical\n    first user messages. Set this when your change specifically tests\n    context-injecting hook behaviour. The engine sets it automatically\n    when the change spec targets hooks or plugins (disable-plugin, or a\n    file change to afk.config.json or a hooks.json manifest).\n";

package/dist/whatif/mde.d.ts

@@ -0,0 +1,9 @@
+export declare const MDE_GATE_THRESHOLD = 0.2;
+export declare const Z_ALPHA = 1.96;
+export declare const Z_POWER = 0.8416;
+export declare function mdeForN(n: number): number;
+export declare function nForMde(mde: number): number;
+export declare function mdePreflightLine(probesPerPrediction: number, targetMde?: number): string;
+export declare function mdeLimitLine(n: number, label: string): string | undefined;
+export declare function isUnderpowered(minProbesPerPrediction: number): boolean;
+export declare function mdeGateRefusedMessage(minProbesPerPrediction: number): string;

package/dist/whatif/report.d.ts

@@ -1,9 +1,10 @@
 import type { ThemePalette } from '../cli/palette.js';
-import type { WhatifReport } from './types.js';
+import type { VerifiedPrediction, WhatifReport } from './types.js';
 export { buildHeadline } from './report.headline.js';
 export declare function renderMarkdown(report: WhatifReport): string;
 export declare function renderTerminal(report: WhatifReport, palette: ThemePalette): string[];
 export declare function standardLimits(opts: {
     verified: boolean;
     judgeExternal: boolean;
+    verifiedPredictions?: VerifiedPrediction[];
 }): string[];

package/dist/whatif/run.d.ts

@@ -4,6 +4,10 @@ export declare class WhatifBudgetError extends Error {
     readonly maxUsd: number;
     constructor(estimateUsd: number, maxUsd: number);
 }
+export declare class WhatifMdeError extends Error {
+    readonly episodesPerArm: number;
+    constructor(minProbesPerPrediction: number);
+}
 export declare function runWhatif(options: WhatifOptions & {
     sessionsDir?: string;
 }, deps: WhatifDeps): Promise<WhatifReport>;

package/dist/whatif/stats.d.ts

@@ -1,8 +1,11 @@
-import type { Prediction, RateComparison, Verdict } from './types.js';
+import type { Prediction, RateComparison, VerifiedPrediction, Verdict } from './types.js';
 export declare function compareRates(baseline: number[], candidate: number[]): RateComparison;
 export declare const EQUIVALENCE_MARGIN = 0.05;
 export declare function verdictFor(pred: Prediction, rates: RateComparison): Verdict;
 export declare function predictionAccuracy(verified: ReadonlyArray<{
     verdict: Verdict;
 }>): number | undefined;
+export declare const CROSS_CHECK_MIN_AGREEMENT = 0.75;
+export declare const CROSS_CHECK_MIN_ITEMS = 5;
+export declare function applyAgreementDowngrade(vp: VerifiedPrediction, mainScores: number[], crossScores: number[]): VerifiedPrediction;
 export declare function agreementRate(a: number[], b: number[]): number;

package/dist/whatif/types.d.ts

@@ -193,6 +193,9 @@ export interface VerifiedPrediction {
     verdict: Verdict;
     scope?: PredictionScope;
     unobservableReason?: string;
+    crossCheckAgreement?: number;
+    crossCheckTooFew?: boolean;
+    verdictReason?: 'judges disagree';
 }
 export interface DiscoveredDifference {
     id: string;

package/dist/whatif/types.orchestration.d.ts

@@ -1,5 +1,5 @@
 import type { AgentRunner, ChangeSpec, Judge } from './types.js';
-export type WhatifStage = 'compile' | 'sandbox' | 'snapshot' | 'predict' | 'episodes' | 'run' | 'judge' | 'discover' | 'report';
+export type WhatifStage = 'compile' | 'sandbox' | 'snapshot' | 'predict' | 'episodes' | 'preflight' | 'run' | 'judge' | 'discover' | 'report';
 export interface WhatifProgress {
     stage: WhatifStage;
     message: string;
@@ -31,6 +31,7 @@ export interface WhatifOptions {
     maxTurns: number;
     episodeTimeoutMs: number;
     keepSandboxes: boolean;
+    force?: boolean;
 }
 export interface WhatifDeps {
     runner: AgentRunner;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-afk",
-  "version": "5.266.2",
+  "version": "5.267.0",
   "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",