agent-afk 5.261.5 → 5.261.7

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";
+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";

package/dist/whatif/index.d.ts

@@ -1,6 +1,6 @@
 export { runWhatif, WhatifBudgetError } from './run.js';
 export { computeStructuralImpact } from './structural.js';
-export type { Change, ChangeKind, ChangeSpec, LaunchSettings, Environment, OperatorContext, ChangeOperator, RequestSnapshot, StructuralImpact, PredictionDirection, Confidence, Prediction, Episode, ToolRequest, EpisodeTrace, RunnerOptions, AgentRunner, EpisodeFeatures, JudgeQuestion, JudgeInput, JudgeResult, Judge, RateComparison, Verdict, VerifiedPrediction, DiscoveredDifference, FeatureDelta, VerifyResult, WhatifReport, WhatifStage, WhatifProgress, CompleteFn, WhatifOptions, WhatifDeps, } from './types.js';
+export type { Change, ChangeKind, ChangeSpec, LaunchSettings, Environment, OperatorContext, ChangeOperator, RequestSnapshot, StructuralImpact, PredictionDirection, Confidence, Prediction, Episode, ToolRequest, EpisodeTrace, RunnerOptions, AgentRunner, EpisodeFeatures, JudgeQuestion, JudgeInput, JudgeResult, Judge, RateComparison, Verdict, VerifiedPrediction, PredictionScope, DiscoveredDifference, FeatureDelta, VerifyResult, WhatifReport, WhatifStage, WhatifProgress, CompleteFn, WhatifOptions, WhatifDeps, } from './types.js';
 export { renderTerminal, renderMarkdown } from './report.js';
 export { createAfkRunner } from './runner/afk-runner.js';
 export { createAnthropicComplete } from './complete.js';

package/dist/whatif/report.d.ts

@@ -1,6 +1,6 @@
 import type { ThemePalette } from '../cli/palette.js';
 import type { WhatifReport } from './types.js';
-export declare function buildHeadline(report: Omit<WhatifReport, 'headline'>): string;
+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: {

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

@@ -0,0 +1,2 @@
+import type { WhatifReport } from './types.js';
+export declare function buildHeadline(report: Omit<WhatifReport, 'headline'>): string;

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

@@ -1,2 +1,8 @@
-import type { VerifyResult } from './types.js';
+import type { VerifyResult, ChangeSpec } from './types.js';
+import type { StructuralImpact } from './types.js';
+export declare function specTargetsHooksOrPlugins(spec: ChangeSpec): boolean;
 export declare function verifyShortfallLimits(v: VerifyResult): string[];
+export declare function hookIsolationLimits(opts: {
+    keepContextHooks: boolean;
+    structural: Pick<StructuralImpact, 'userMessageDiff'>;
+}): string[];

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

@@ -0,0 +1,7 @@
+import type { Episode, EpisodeTrace, Prediction, VerifiedPrediction } from './types.js';
+type Arm = 'baseline' | 'candidate';
+export type JudgeResults = Map<string, Record<string, number>>;
+export declare function traceKey(trace: Pick<EpisodeTrace, 'episodeId' | 'env' | 'sample'>): string;
+export declare function scoresForQuestion(qid: string, env: Arm, goodTraces: EpisodeTrace[], judgeResults: JudgeResults, include?: (episodeId: string) => boolean): number[];
+export declare function scorePrediction(prediction: Prediction, episodes: Episode[], goodTraces: EpisodeTrace[], judgeResults: JudgeResults): VerifiedPrediction;
+export {};

package/dist/whatif/types.d.ts

@@ -175,10 +175,19 @@ export interface RateComparison {
     };
 }
 export type Verdict = 'confirmed' | 'refuted' | 'unclear';
+export interface PredictionScope {
+    episodes: {
+        baseline: string[];
+        candidate: string[];
+    };
+    targetedEpisodes: number;
+    background?: RateComparison;
+}
 export interface VerifiedPrediction {
     prediction: Prediction;
     rates: RateComparison;
     verdict: Verdict;
+    scope?: PredictionScope;
 }
 export interface DiscoveredDifference {
     id: string;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-afk",
-  "version": "5.261.5",
+  "version": "5.261.7",
   "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",