agent-afk 5.271.3 → 5.272.1

package/dist/whatif/args.d.ts

@@ -15,6 +15,7 @@ export interface WhatifFlagOptions {
     keepSandboxes: boolean;
     probes?: number;
     maxPredictions?: number;
+    noBaselineSample?: boolean;
 }
 export interface ParsedWhatifArgs {
     flagChanges: Change[];

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  --probes <n>                 Synthetic probe episodes per prediction (1\u201312; default 6)\n  --max-predictions <n>        Max predictions to retain (1\u20138; default 3 when probes>2, else 8)\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";
+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  --probes <n>                 Synthetic probe episodes per prediction (1\u201312; default 6)\n  --max-predictions <n>        Max predictions to retain (1\u20138; default 3 when probes>2, else 8)\n  --keep-sandboxes             Keep sandbox directories after run\n  --no-baseline-sample         Skip the baseline-sample preflight; use analyst estimate instead (#2511)\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/baseline-sample.d.ts

@@ -0,0 +1,38 @@
+import type { AgentRunner, Episode, Judge, Prediction, RunnerOptions, Environment, WhatifProgress } from './types.js';
+export declare const BASELINE_SAMPLE_K = 3;
+export interface PredictionBaselineSample {
+    predictionId: string;
+    probeRates: number[];
+    mean: number;
+    optimisticHeadroom: number;
+    mde: number;
+    tripped: boolean;
+}
+export interface BaselineSampleResult {
+    perPrediction: PredictionBaselineSample[];
+    anyTripped: boolean;
+    agentCostUsd: number;
+    firstTrip?: {
+        predictionId: string;
+        fullRunProbes: number;
+        message: string;
+    };
+}
+export declare function estimateBaselineSampleCost(input: {
+    predictions: Prediction[];
+    agentModel: string;
+    analystModel: string;
+    systemTokensBaseline: number;
+    judgeExternal: boolean;
+}): number;
+export declare function runBaselineSample(input: {
+    predictions: Prediction[];
+    episodes: Episode[];
+    baseline: Environment;
+    runner: AgentRunner;
+    judge: Judge;
+    runnerOpts: RunnerOptions;
+    warnOnly: boolean;
+    signal?: AbortSignal;
+    onProgress?: (p: WhatifProgress) => void;
+}): Promise<BaselineSampleResult>;

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

@@ -0,0 +1,66 @@
+import { type BaselineSampleResult } from './baseline-sample.js';
+export interface PreflightInput {
+    episodesPerArm: number;
+    minProbesPerPrediction: number;
+    syntheticPerArm: number;
+    predictionCount: number;
+    predictions: import('./types.js').Prediction[];
+    force: boolean;
+    samples: number;
+    agentModel: string;
+    analystModel: string;
+    systemTokens: {
+        baseline: number;
+        candidate: number;
+    };
+    judgeExternal: boolean;
+    analystCostUsd: number;
+    maxUsd: number;
+    baselineSampleCostUsd: number;
+    measuredHeadroomPending?: boolean;
+    onProgress: ((p: {
+        stage: 'preflight';
+        message: string;
+    }) => void) | undefined;
+}
+export declare function runPreflightChecks(input: PreflightInput): void;
+export declare function resolveMinProbesPerPrediction(predictions: import('./types.js').Prediction[], episodes: import('./types.js').Episode[]): number;
+export interface BaselineSamplePhaseInput {
+    noBaselineSample: boolean;
+    predictions: import('./types.js').Prediction[];
+    episodes: import('./types.js').Episode[];
+    baseline: import('./types.js').Environment;
+    runner: import('./types.js').AgentRunner;
+    judge: import('./types.js').Judge;
+    episodeTimeoutMs: number;
+    maxTurns: number;
+    signal?: AbortSignal;
+    onProgress?: (p: import('./types.js').WhatifProgress) => void;
+    closeJudges: () => Promise<void>;
+    runDir: string;
+}
+export declare function runBaselineSamplePhase(input: BaselineSamplePhaseInput): Promise<BaselineSampleResult | undefined>;
+export interface VerifyPreflightInput {
+    episodes: import('./types.js').Episode[];
+    predictions: import('./types.js').Prediction[];
+    structural: import('./types.js').StructuralImpact;
+    force: boolean;
+    samples: number;
+    agentModel: string;
+    analystModel: string;
+    judgeExternal: boolean;
+    analystCostUsd: number;
+    maxUsd: number;
+    noBaselineSample: boolean;
+    onProgress: ((p: {
+        stage: 'preflight';
+        message: string;
+    }) => void) | undefined;
+    closeJudges: () => Promise<void>;
+}
+export declare function runVerifyPreflight(input: VerifyPreflightInput): Promise<{
+    episodesPerArm: number;
+    minProbesPerPrediction: number;
+    noBaselineSample: boolean;
+    baselineSampleCostUsd: number;
+}>;

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

@@ -16,5 +16,6 @@ export interface VerifiedReportArgs {
     resolvedJudge: Judge;
     autoKeepContextHooks: boolean;
     judgeResults: JudgeResults;
+    baselineSamplePerPrediction?: import('./baseline-sample.js').PredictionBaselineSample[];
 }
 export declare function buildAndPersistVerifiedReport(args: VerifiedReportArgs): Promise<WhatifReport>;

package/dist/whatif/types.d.ts

@@ -226,6 +226,7 @@ export interface VerifyResult {
     judgeFailures?: number;
     failedEpisodeRecords?: import('./run.failures.js').FailedEpisodeRecord[];
     armImbalance?: import('./run.failures.js').ArmImbalance;
+    baselineSample?: import('./baseline-sample.js').PredictionBaselineSample[];
 }
 export interface WhatifReport {
     spec: ChangeSpec;

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

@@ -34,6 +34,7 @@ export interface WhatifOptions {
     force?: boolean;
     probes?: number;
     maxPredictions?: number;
+    noBaselineSample?: boolean;
 }
 export interface WhatifDeps {
     runner: AgentRunner;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agent-afk",
-  "version": "5.271.3",
+  "version": "5.272.1",
   "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",