@caupulican/pi-adaptative 0.80.29 → 0.80.31

package/dist/core/improvement-loop.d.ts

@@ -0,0 +1,296 @@
+import { Type } from "typebox";
+import { type ToolDefinition } from "./extensions/types.ts";
+export type MetricDirection = "lower" | "higher";
+export type ImprovementDecision = "keep" | "discard" | "retry" | "blocked";
+export type ImprovementDecisionReason = "baseline" | "metric_improved" | "checks_failed" | "metric_missing" | "metric_invalid" | "not_better_than_best" | "below_min_delta" | "below_min_relative_delta" | "insufficient_noise_evidence" | "below_confidence";
+export type ConfidenceMode = "none" | "mad";
+export type LowConfidenceAction = "retry" | "discard";
+export interface MetricComparison {
+    current: number;
+    best: number | null;
+    direction: MetricDirection;
+    delta: number;
+    relativeDelta: number | null;
+    improved: boolean;
+}
+export interface ConfidenceResult {
+    mode: ConfidenceMode;
+    value: number | null;
+    noiseFloor: number | null;
+    sampleCount: number;
+}
+export interface ImprovementDecisionInput {
+    currentMetric: number | null | undefined;
+    bestMetric?: number | null;
+    baselineMetric?: number | null;
+    direction?: MetricDirection;
+    checksPass?: boolean | null;
+    minDelta?: number;
+    minRelativeDelta?: number;
+    confidenceMode?: ConfidenceMode;
+    minConfidence?: number;
+    lowConfidenceAction?: LowConfidenceAction;
+    noiseMetrics?: number[];
+}
+export interface ImprovementDecisionResult {
+    decision: ImprovementDecision;
+    reason: ImprovementDecisionReason;
+    comparison: MetricComparison | null;
+    confidence: ConfidenceResult;
+}
+export interface ImprovementRunRecord {
+    runId: string | number;
+    objective?: string;
+    hypothesis?: string;
+    metricName: string;
+    metricUnit?: string;
+    direction: MetricDirection;
+    metric: number | null;
+    secondaryMetrics?: Record<string, number>;
+    checksPass: boolean | null;
+    decision: ImprovementDecision;
+    reason: ImprovementDecisionReason;
+    confidence?: number | null;
+    changedFiles?: string[];
+    evidenceRef?: string;
+    nextHint?: string;
+    timestamp: number;
+}
+export interface GitStatusEntry {
+    index: string;
+    workingTree: string;
+    path: string;
+    origPath?: string;
+}
+export interface OwnedDiscardPlanInput {
+    beforeStatus: string | GitStatusEntry[];
+    afterStatus: string | GitStatusEntry[];
+    ownedPaths: string[];
+    preservePaths?: string[];
+}
+export interface OwnedDiscardPlan {
+    revertPaths: string[];
+    preservePaths: string[];
+    protectedUserDirtyPaths: string[];
+    unownedChangedPaths: string[];
+    canDiscardOwnedChanges: boolean;
+}
+export interface ImprovementLoopConfig {
+    loopId: string;
+    objective: string;
+    metricName: string;
+    metricUnit?: string;
+    direction: MetricDirection;
+    minDelta: number;
+    minRelativeDelta: number;
+    confidenceMode: ConfidenceMode;
+    minConfidence: number;
+    lowConfidenceAction: LowConfidenceAction;
+    createdAt: number;
+    cwd: string;
+}
+export interface ImprovementSandboxRecord {
+    sandboxId: string;
+    status: "active" | "cleaned";
+    repoPath: string;
+    worktreePath: string;
+    baseRef: string;
+    createdAt: number;
+    cleanedAt?: number;
+    exportedAt?: number;
+    patchPath?: string;
+    patchBytes?: number;
+    reason?: string;
+}
+export interface ImprovementLoopState {
+    config: ImprovementLoopConfig;
+    runs: ImprovementRunRecord[];
+    sandboxes: ImprovementSandboxRecord[];
+    activeSandbox: ImprovementSandboxRecord | null;
+    baselineMetric: number | null;
+    bestMetric: number | null;
+    bestRunId: string | number | null;
+    lastDecision: ImprovementDecisionResult | null;
+    logPath: string;
+}
+export interface ImprovementLoopPaths {
+    rootDir: string;
+    workspaceDir: string;
+    logPath: string;
+    sandboxDir: string;
+    artifactDir: string;
+    workspaceKey: string;
+    loopId: string;
+}
+export interface ImprovementLoopInitInput {
+    cwd: string;
+    objective: string;
+    metricName: string;
+    metricUnit?: string;
+    direction?: MetricDirection;
+    loopId?: string;
+    minDelta?: number;
+    minRelativeDelta?: number;
+    confidenceMode?: ConfidenceMode;
+    minConfidence?: number;
+    lowConfidenceAction?: LowConfidenceAction;
+    agentDir?: string;
+    reset?: boolean;
+}
+export interface ImprovementLoopRecordInput {
+    cwd: string;
+    loopId?: string;
+    runId?: string | number;
+    metric: number | null | undefined;
+    secondaryMetrics?: Record<string, number>;
+    checksPass?: boolean | null;
+    hypothesis?: string;
+    changedFiles?: string[];
+    evidenceRef?: string;
+    nextHint?: string;
+    agentDir?: string;
+}
+export interface ImprovementLoopStatusInput {
+    cwd: string;
+    loopId?: string;
+    agentDir?: string;
+}
+export interface ImprovementSandboxCreateInput extends ImprovementLoopStatusInput {
+    exec: ImprovementLoopExec;
+    baseRef?: string;
+    allowDirtyRepo?: boolean;
+    sandboxId?: string;
+    signal?: AbortSignal;
+}
+export interface ImprovementSandboxCleanupInput extends ImprovementLoopStatusInput {
+    exec: ImprovementLoopExec;
+    sandboxId?: string;
+    reason?: string;
+    signal?: AbortSignal;
+}
+export interface ImprovementSandboxExportInput extends ImprovementLoopStatusInput {
+    exec: ImprovementLoopExec;
+    sandboxId?: string;
+    allowEmptyPatch?: boolean;
+    signal?: AbortSignal;
+}
+export type ImprovementLoopToolAction = "init" | "status" | "record" | "measure" | "sandbox_create" | "sandbox_export" | "sandbox_cleanup";
+export interface ImprovementMeasurementCommandResult {
+    stdout: string;
+    stderr: string;
+    code: number;
+    killed: boolean;
+    stdoutTruncated?: boolean;
+    stderrTruncated?: boolean;
+}
+export interface ImprovementMeasurementResult {
+    command: string;
+    exitCode: number;
+    timedOut: boolean;
+    durationMs: number;
+    stdout: string;
+    stderr: string;
+    stdoutTruncated: boolean;
+    stderrTruncated: boolean;
+    parsedMetrics: Record<string, number>;
+    primaryMetric: number | null;
+    checksCommand?: string;
+    checksExitCode?: number;
+    checksTimedOut?: boolean;
+    checksPass: boolean;
+    checksStdout?: string;
+    checksStderr?: string;
+}
+export type ImprovementLoopExec = (command: string, args: string[], options?: {
+    cwd?: string;
+    timeout?: number;
+    signal?: AbortSignal;
+    maxBuffer?: number;
+}) => Promise<ImprovementMeasurementCommandResult>;
+export interface ImprovementLoopToolDetails {
+    action: ImprovementLoopToolAction;
+    state: ImprovementLoopState | null;
+    decision?: ImprovementDecisionResult;
+    measurement?: ImprovementMeasurementResult;
+    sandbox?: ImprovementSandboxRecord;
+    logPath: string;
+}
+export declare function parseMetricLines(output: string): Map<string, number>;
+export declare function metricMapFromOutput(output: string): Record<string, number>;
+export declare function selectPrimaryMetric(metrics: Map<string, number> | Record<string, number>, metricName: string): number | null;
+export declare function compareMetric(current: number, best: number | null | undefined, direction?: MetricDirection): MetricComparison;
+export declare function median(values: number[]): number | null;
+export declare function medianAbsoluteDeviation(values: number[]): number | null;
+export declare function computeMadConfidence(improvementDelta: number, samples: number[]): ConfidenceResult;
+export declare function decideImprovement(input: ImprovementDecisionInput): ImprovementDecisionResult;
+export declare function serializeRunRecord(record: ImprovementRunRecord): string;
+export declare function appendRunRecord(filePath: string, record: ImprovementRunRecord): Promise<void>;
+export declare function readRunRecords(filePath: string): Promise<ImprovementRunRecord[]>;
+export declare function improvementLoopPaths(input: ImprovementLoopStatusInput): ImprovementLoopPaths;
+export declare function initImprovementLoop(input: ImprovementLoopInitInput): Promise<ImprovementLoopState>;
+export declare function readImprovementLoopState(input: ImprovementLoopStatusInput): Promise<ImprovementLoopState | null>;
+export declare function createImprovementSandbox(input: ImprovementSandboxCreateInput): Promise<ImprovementLoopState>;
+export declare function exportImprovementSandboxPatch(input: ImprovementSandboxExportInput): Promise<ImprovementLoopState>;
+export declare function cleanupImprovementSandbox(input: ImprovementSandboxCleanupInput): Promise<ImprovementLoopState>;
+export declare function recordImprovementRun(input: ImprovementLoopRecordInput): Promise<ImprovementLoopState>;
+export declare function runImprovementMeasurement(input: {
+    exec: ImprovementLoopExec;
+    cwd: string;
+    command: string;
+    metricName: string;
+    checksCommand?: string;
+    timeoutSeconds?: number;
+    checksTimeoutSeconds?: number;
+    maxOutputBytes?: number;
+    signal?: AbortSignal;
+}): Promise<ImprovementMeasurementResult>;
+export declare function parseGitPorcelainStatus(status: string): GitStatusEntry[];
+export declare function planOwnedDiscard(input: OwnedDiscardPlanInput): OwnedDiscardPlan;
+declare const DecisionToolParams: Type.TObject<{
+    currentMetric: Type.TNumber;
+    bestMetric: Type.TOptional<Type.TNumber>;
+    direction: Type.TUnsafe<"higher" | "lower">;
+    checksPass: Type.TOptional<Type.TBoolean>;
+    minDelta: Type.TOptional<Type.TNumber>;
+    minRelativeDelta: Type.TOptional<Type.TNumber>;
+    confidenceMode: Type.TOptional<Type.TUnsafe<"mad" | "none">>;
+    minConfidence: Type.TOptional<Type.TNumber>;
+    lowConfidenceAction: Type.TOptional<Type.TUnsafe<"discard" | "retry">>;
+    noiseMetrics: Type.TOptional<Type.TArray<Type.TNumber>>;
+}>;
+export declare function createImprovementDecisionTool(): ToolDefinition<typeof DecisionToolParams, ImprovementDecisionResult>;
+declare const LoopToolParams: Type.TObject<{
+    action: Type.TUnsafe<"init" | "measure" | "record" | "sandbox_cleanup" | "sandbox_create" | "sandbox_export" | "status">;
+    loopId: Type.TOptional<Type.TString>;
+    objective: Type.TOptional<Type.TString>;
+    metricName: Type.TOptional<Type.TString>;
+    metricUnit: Type.TOptional<Type.TString>;
+    direction: Type.TOptional<Type.TUnsafe<"higher" | "lower">>;
+    reset: Type.TOptional<Type.TBoolean>;
+    minDelta: Type.TOptional<Type.TNumber>;
+    minRelativeDelta: Type.TOptional<Type.TNumber>;
+    confidenceMode: Type.TOptional<Type.TUnsafe<"mad" | "none">>;
+    minConfidence: Type.TOptional<Type.TNumber>;
+    lowConfidenceAction: Type.TOptional<Type.TUnsafe<"discard" | "retry">>;
+    currentMetric: Type.TOptional<Type.TNumber>;
+    checksPass: Type.TOptional<Type.TBoolean>;
+    hypothesis: Type.TOptional<Type.TString>;
+    secondaryMetrics: Type.TOptional<Type.TObject<{}>>;
+    changedFiles: Type.TOptional<Type.TArray<Type.TString>>;
+    evidenceRef: Type.TOptional<Type.TString>;
+    nextHint: Type.TOptional<Type.TString>;
+    command: Type.TOptional<Type.TString>;
+    checksCommand: Type.TOptional<Type.TString>;
+    timeoutSeconds: Type.TOptional<Type.TNumber>;
+    checksTimeoutSeconds: Type.TOptional<Type.TNumber>;
+    maxOutputBytes: Type.TOptional<Type.TNumber>;
+    useSandbox: Type.TOptional<Type.TBoolean>;
+    sandboxId: Type.TOptional<Type.TString>;
+    baseRef: Type.TOptional<Type.TString>;
+    allowDirtyRepo: Type.TOptional<Type.TBoolean>;
+    cleanupReason: Type.TOptional<Type.TString>;
+    allowEmptyPatch: Type.TOptional<Type.TBoolean>;
+}>;
+export declare function createImprovementLoopTool(exec?: ImprovementLoopExec): ToolDefinition<typeof LoopToolParams, ImprovementLoopToolDetails>;
+export {};
+//# sourceMappingURL=improvement-loop.d.ts.map

package/dist/core/improvement-loop.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"improvement-loop.d.ts","sourceRoot":"","sources":["../../src/core/improvement-loop.ts"],"names":[],"mappings":"AAIA,OAAO,EAAE,IAAI,EAAE,MAAM,SAAS,CAAC;AAE/B,OAAO,EAAc,KAAK,cAAc,EAAE,MAAM,uBAAuB,CAAC;AAExE,MAAM,MAAM,eAAe,GAAG,OAAO,GAAG,QAAQ,CAAC;AACjD,MAAM,MAAM,mBAAmB,GAAG,MAAM,GAAG,SAAS,GAAG,OAAO,GAAG,SAAS,CAAC;AAC3E,MAAM,MAAM,yBAAyB,GAClC,UAAU,GACV,iBAAiB,GACjB,eAAe,GACf,gBAAgB,GAChB,gBAAgB,GAChB,sBAAsB,GACtB,iBAAiB,GACjB,0BAA0B,GAC1B,6BAA6B,GAC7B,kBAAkB,CAAC;AACtB,MAAM,MAAM,cAAc,GAAG,MAAM,GAAG,KAAK,CAAC;AAC5C,MAAM,MAAM,mBAAmB,GAAG,OAAO,GAAG,SAAS,CAAC;AAEtD,MAAM,WAAW,gBAAgB;IAChC,OAAO,EAAE,MAAM,CAAC;IAChB,IAAI,EAAE,MAAM,GAAG,IAAI,CAAC;IACpB,SAAS,EAAE,eAAe,CAAC;IAC3B,KAAK,EAAE,MAAM,CAAC;IACd,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,QAAQ,EAAE,OAAO,CAAC;CAClB;AAED,MAAM,WAAW,gBAAgB;IAChC,IAAI,EAAE,cAAc,CAAC;IACrB,KAAK,EAAE,MAAM,GAAG,IAAI,CAAC;IACrB,UAAU,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1B,WAAW,EAAE,MAAM,CAAC;CACpB;AAED,MAAM,WAAW,wBAAwB;IACxC,aAAa,EAAE,MAAM,GAAG,IAAI,GAAG,SAAS,CAAC;IACzC,UAAU,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,cAAc,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAC/B,SAAS,CAAC,EAAE,eAAe,CAAC;IAC5B,UAAU,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC;IAC5B,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,cAAc,CAAC,EAAE,cAAc,CAAC;IAChC,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,mBAAmB,CAAC,EAAE,mBAAmB,CAAC;IAC1C,YAAY,CAAC,EAAE,MAAM,EAAE,CAAC;CACxB;AAED,MAAM,WAAW,yBAAyB;IACzC,QAAQ,EAAE,mBAAmB,CAAC;IAC9B,MAAM,EAAE,yBAAyB,CAAC;IAClC,UAAU,EAAE,gBAAgB,GAAG,IAAI,CAAC;IACpC,UAAU,EAAE,gBAAgB,CAAC;CAC7B;AAED,MAAM,WAAW,oBAAoB;IACpC,KAAK,EAAE,MAAM,GAAG,MAAM,CAAC;IACvB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,SAAS,EAAE,eAAe,CAAC;IAC3B,MAAM,EAAE,MAAM,GAAG,IAAI,CAAC;IACtB,gBAAgB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC1C,UAAU,EAAE,OAAO,GAAG,IAAI,CAAC;IAC3B,QAAQ,EAAE,mBAAmB,CAAC;IAC9B,MAAM,EAAE,yBAAyB,CAAC;IAClC,UAAU,CAAC,EAAE,MAAM,GAAG,IAAI,CAAC;IAC3B,YAAY,CAAC,EAAE,MAAM,EAAE,CAAC;IACxB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,SAAS,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,cAAc;IAC9B,KAAK,EAAE,MAAM,CAAC;IACd,WAAW,EAAE,MAAM,CAAC;IACpB,IAAI,EAAE,MAAM,CAAC;IACb,QAAQ,CAAC,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,qBAAqB;IACrC,YAAY,EAAE,MAAM,GAAG,cAAc,EAAE,CAAC;IACxC,WAAW,EAAE,MAAM,GAAG,cAAc,EAAE,CAAC;IACvC,UAAU,EAAE,MAAM,EAAE,CAAC;IACrB,aAAa,CAAC,EAAE,MAAM,EAAE,CAAC;CACzB;AAED,MAAM,WAAW,gBAAgB;IAChC,WAAW,EAAE,MAAM,EAAE,CAAC;IACtB,aAAa,EAAE,MAAM,EAAE,CAAC;IACxB,uBAAuB,EAAE,MAAM,EAAE,CAAC;IAClC,mBAAmB,EAAE,MAAM,EAAE,CAAC;IAC9B,sBAAsB,EAAE,OAAO,CAAC;CAChC;AAED,MAAM,WAAW,qBAAqB;IACrC,MAAM,EAAE,MAAM,CAAC;IACf,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,SAAS,EAAE,eAAe,CAAC;IAC3B,QAAQ,EAAE,MAAM,CAAC;IACjB,gBAAgB,EAAE,MAAM,CAAC;IACzB,cAAc,EAAE,cAAc,CAAC;IAC/B,aAAa,EAAE,MAAM,CAAC;IACtB,mBAAmB,EAAE,mBAAmB,CAAC;IACzC,SAAS,EAAE,MAAM,CAAC;IAClB,GAAG,EAAE,MAAM,CAAC;CACZ;AAED,MAAM,WAAW,wBAAwB;IACxC,SAAS,EAAE,MAAM,CAAC;IAClB,MAAM,EAAE,QAAQ,GAAG,SAAS,CAAC;IAC7B,QAAQ,EAAE,MAAM,CAAC;IACjB,YAAY,EAAE,MAAM,CAAC;IACrB,OAAO,EAAE,MAAM,CAAC;IAChB,SAAS,EAAE,MAAM,CAAC;IAClB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,MAAM,CAAC,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,oBAAoB;IACpC,MAAM,EAAE,qBAAqB,CAAC;IAC9B,IAAI,EAAE,oBAAoB,EAAE,CAAC;IAC7B,SAAS,EAAE,wBAAwB,EAAE,CAAC;IACtC,aAAa,EAAE,wBAAwB,GAAG,IAAI,CAAC;IAC/C,cAAc,EAAE,MAAM,GAAG,IAAI,CAAC;IAC9B,UAAU,EAAE,MAAM,GAAG,IAAI,CAAC;IAC1B,SAAS,EAAE,MAAM,GAAG,MAAM,GAAG,IAAI,CAAC;IAClC,YAAY,EAAE,yBAAyB,GAAG,IAAI,CAAC;IAC/C,OAAO,EAAE,MAAM,CAAC;CAChB;AAED,MAAM,WAAW,oBAAoB;IACpC,OAAO,EAAE,MAAM,CAAC;IAChB,YAAY,EAAE,MAAM,CAAC;IACrB,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,WAAW,EAAE,MAAM,CAAC;IACpB,YAAY,EAAE,MAAM,CAAC;IACrB,MAAM,EAAE,MAAM,CAAC;CACf;AAED,MAAM,WAAW,wBAAwB;IACxC,GAAG,EAAE,MAAM,CAAC;IACZ,SAAS,EAAE,MAAM,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,SAAS,CAAC,EAAE,eAAe,CAAC;IAC5B,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,gBAAgB,CAAC,EAAE,MAAM,CAAC;IAC1B,cAAc,CAAC,EAAE,cAAc,CAAC;IAChC,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,mBAAmB,CAAC,EAAE,mBAAmB,CAAC;IAC1C,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,KAAK,CAAC,EAAE,OAAO,CAAC;CAChB;AAED,MAAM,WAAW,0BAA0B;IAC1C,GAAG,EAAE,MAAM,CAAC;IACZ,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,KAAK,CAAC,EAAE,MAAM,GAAG,MAAM,CAAC;IACxB,MAAM,EAAE,MAAM,GAAG,IAAI,GAAG,SAAS,CAAC;IAClC,gBAAgB,CAAC,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IAC1C,UAAU,CAAC,EAAE,OAAO,GAAG,IAAI,CAAC;IAC5B,UAAU,CAAC,EAAE,MAAM,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,EAAE,CAAC;IACxB,WAAW,CAAC,EAAE,MAAM,CAAC;IACrB,QAAQ,CAAC,EAAE,MAAM,CAAC;IAClB,QAAQ,CAAC,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,0BAA0B;IAC1C,GAAG,EAAE,MAAM,CAAC;IACZ,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,QAAQ,CAAC,EAAE,MAAM,CAAC;CAClB;AAED,MAAM,WAAW,6BAA8B,SAAQ,0BAA0B;IAChF,IAAI,EAAE,mBAAmB,CAAC;IAC1B,OAAO,CAAC,EAAE,MAAM,CAAC;IACjB,cAAc,CAAC,EAAE,OAAO,CAAC;IACzB,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAED,MAAM,WAAW,8BAA+B,SAAQ,0BAA0B;IACjF,IAAI,EAAE,mBAAmB,CAAC;IAC1B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,MAAM,CAAC,EAAE,MAAM,CAAC;IAChB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAED,MAAM,WAAW,6BAA8B,SAAQ,0BAA0B;IAChF,IAAI,EAAE,mBAAmB,CAAC;IAC1B,SAAS,CAAC,EAAE,MAAM,CAAC;IACnB,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB;AAOD,MAAM,MAAM,yBAAyB,GAClC,MAAM,GACN,QAAQ,GACR,QAAQ,GACR,SAAS,GACT,gBAAgB,GAChB,gBAAgB,GAChB,iBAAiB,CAAC;AAErB,MAAM,WAAW,mCAAmC;IACnD,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,MAAM,CAAC;IACf,IAAI,EAAE,MAAM,CAAC;IACb,MAAM,EAAE,OAAO,CAAC;IAChB,eAAe,CAAC,EAAE,OAAO,CAAC;IAC1B,eAAe,CAAC,EAAE,OAAO,CAAC;CAC1B;AAED,MAAM,WAAW,4BAA4B;IAC5C,OAAO,EAAE,MAAM,CAAC;IAChB,QAAQ,EAAE,MAAM,CAAC;IACjB,QAAQ,EAAE,OAAO,CAAC;IAClB,UAAU,EAAE,MAAM,CAAC;IACnB,MAAM,EAAE,MAAM,CAAC;IACf,MAAM,EAAE,MAAM,CAAC;IACf,eAAe,EAAE,OAAO,CAAC;IACzB,eAAe,EAAE,OAAO,CAAC;IACzB,aAAa,EAAE,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAAC;IACtC,aAAa,EAAE,MAAM,GAAG,IAAI,CAAC;IAC7B,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,cAAc,CAAC,EAAE,OAAO,CAAC;IACzB,UAAU,EAAE,OAAO,CAAC;IACpB,YAAY,CAAC,EAAE,MAAM,CAAC;IACtB,YAAY,CAAC,EAAE,MAAM,CAAC;CACtB;AAED,MAAM,MAAM,mBAAmB,GAAG,CACjC,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,MAAM,EAAE,EACd,OAAO,CAAC,EAAE;IAAE,GAAG,CAAC,EAAE,MAAM,CAAC;IAAC,OAAO,CAAC,EAAE,MAAM,CAAC;IAAC,MAAM,CAAC,EAAE,WAAW,CAAC;IAAC,SAAS,CAAC,EAAE,MAAM,CAAA;CAAE,KAClF,OAAO,CAAC,mCAAmC,CAAC,CAAC;AAElD,MAAM,WAAW,0BAA0B;IAC1C,MAAM,EAAE,yBAAyB,CAAC;IAClC,KAAK,EAAE,oBAAoB,GAAG,IAAI,CAAC;IACnC,QAAQ,CAAC,EAAE,yBAAyB,CAAC;IACrC,WAAW,CAAC,EAAE,4BAA4B,CAAC;IAC3C,OAAO,CAAC,EAAE,wBAAwB,CAAC;IACnC,OAAO,EAAE,MAAM,CAAC;CAChB;AAqBD,wBAAgB,gBAAgB,CAAC,MAAM,EAAE,MAAM,GAAG,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,CAgBpE;AAED,wBAAgB,mBAAmB,CAAC,MAAM,EAAE,MAAM,GAAG,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,CAE1E;AAED,wBAAgB,mBAAmB,CAClC,OAAO,EAAE,GAAG,CAAC,MAAM,EAAE,MAAM,CAAC,GAAG,MAAM,CAAC,MAAM,EAAE,MAAM,CAAC,EACrD,UAAU,EAAE,MAAM,GAChB,MAAM,GAAG,IAAI,CAGf;AAED,wBAAgB,aAAa,CAC5B,OAAO,EAAE,MAAM,EACf,IAAI,EAAE,MAAM,GAAG,IAAI,GAAG,SAAS,EAC/B,SAAS,GAAE,eAAyB,GAClC,gBAAgB,CAuBlB;AAED,wBAAgB,MAAM,CAAC,MAAM,EAAE,MAAM,EAAE,GAAG,MAAM,GAAG,IAAI,CAKtD;AAED,wBAAgB,uBAAuB,CAAC,MAAM,EAAE,MAAM,EAAE,GAAG,MAAM,GAAG,IAAI,CAIvE;AAED,wBAAgB,oBAAoB,CAAC,gBAAgB,EAAE,MAAM,EAAE,OAAO,EAAE,MAAM,EAAE,GAAG,gBAAgB,CAelG;AAED,wBAAgB,iBAAiB,CAAC,KAAK,EAAE,wBAAwB,GAAG,yBAAyB,CA8D5F;AAED,wBAAgB,kBAAkB,CAAC,MAAM,EAAE,oBAAoB,GAAG,MAAM,CAEvE;AAED,wBAAsB,eAAe,CAAC,QAAQ,EAAE,MAAM,EAAE,MAAM,EAAE,oBAAoB,GAAG,OAAO,CAAC,IAAI,CAAC,CAGnG;AAED,wBAAsB,cAAc,CAAC,QAAQ,EAAE,MAAM,GAAG,OAAO,CAAC,oBAAoB,EAAE,CAAC,CAetF;AAcD,wBAAgB,oBAAoB,CAAC,KAAK,EAAE,0BAA0B,GAAG,oBAAoB,CAc5F;AAkGD,wBAAsB,mBAAmB,CAAC,KAAK,EAAE,wBAAwB,GAAG,OAAO,CAAC,oBAAoB,CAAC,CAcxG;AAED,wBAAsB,wBAAwB,CAC7C,KAAK,EAAE,0BAA0B,GAC/B,OAAO,CAAC,oBAAoB,GAAG,IAAI,CAAC,CAGtC;AA6BD,wBAAsB,wBAAwB,CAAC,KAAK,EAAE,6BAA6B,GAAG,OAAO,CAAC,oBAAoB,CAAC,CAoClH;AAED,wBAAsB,6BAA6B,CAClD,KAAK,EAAE,6BAA6B,GAClC,OAAO,CAAC,oBAAoB,CAAC,CAoB/B;AAED,wBAAsB,yBAAyB,CAAC,KAAK,EAAE,8BAA8B,GAAG,OAAO,CAAC,oBAAoB,CAAC,CAsBpH;AAED,wBAAsB,oBAAoB,CAAC,KAAK,EAAE,0BAA0B,GAAG,OAAO,CAAC,oBAAoB,CAAC,CA0C3G;AAUD,wBAAsB,yBAAyB,CAAC,KAAK,EAAE;IACtD,IAAI,EAAE,mBAAmB,CAAC;IAC1B,GAAG,EAAE,MAAM,CAAC;IACZ,OAAO,EAAE,MAAM,CAAC;IAChB,UAAU,EAAE,MAAM,CAAC;IACnB,aAAa,CAAC,EAAE,MAAM,CAAC;IACvB,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,oBAAoB,CAAC,EAAE,MAAM,CAAC;IAC9B,cAAc,CAAC,EAAE,MAAM,CAAC;IACxB,MAAM,CAAC,EAAE,WAAW,CAAC;CACrB,GAAG,OAAO,CAAC,4BAA4B,CAAC,CAuDxC;AAED,wBAAgB,uBAAuB,CAAC,MAAM,EAAE,MAAM,GAAG,cAAc,EAAE,CAexE;AAiCD,wBAAgB,gBAAgB,CAAC,KAAK,EAAE,qBAAqB,GAAG,gBAAgB,CAiC/E;AAED,QAAA,MAAM,kBAAkB;;;;;;;;;;;EAkCvB,CAAC;AAEF,wBAAgB,6BAA6B,IAAI,cAAc,CAAC,OAAO,kBAAkB,EAAE,yBAAyB,CAAC,CA6BpH;AAED,QAAA,MAAM,cAAc;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;EA0GnB,CAAC;AAwBF,wBAAgB,yBAAyB,CACxC,IAAI,CAAC,EAAE,mBAAmB,GACxB,cAAc,CAAC,OAAO,cAAc,EAAE,0BAA0B,CAAC,CA8LnE","sourcesContent":["import { createHash } from \"node:crypto\";\nimport { mkdir, readFile, writeFile } from \"node:fs/promises\";\nimport { dirname, join } from \"node:path\";\nimport { StringEnum } from \"@caupulican/pi-ai\";\nimport { Type } from \"typebox\";\nimport { getAgentDir } from \"../config.ts\";\nimport { defineTool, type ToolDefinition } from \"./extensions/types.ts\";\n\nexport type MetricDirection = \"lower\" | \"higher\";\nexport type ImprovementDecision = \"keep\" | \"discard\" | \"retry\" | \"blocked\";\nexport type ImprovementDecisionReason =\n\t| \"baseline\"\n\t| \"metric_improved\"\n\t| \"checks_failed\"\n\t| \"metric_missing\"\n\t| \"metric_invalid\"\n\t| \"not_better_than_best\"\n\t| \"below_min_delta\"\n\t| \"below_min_relative_delta\"\n\t| \"insufficient_noise_evidence\"\n\t| \"below_confidence\";\nexport type ConfidenceMode = \"none\" | \"mad\";\nexport type LowConfidenceAction = \"retry\" | \"discard\";\n\nexport interface MetricComparison {\n\tcurrent: number;\n\tbest: number | null;\n\tdirection: MetricDirection;\n\tdelta: number;\n\trelativeDelta: number | null;\n\timproved: boolean;\n}\n\nexport interface ConfidenceResult {\n\tmode: ConfidenceMode;\n\tvalue: number | null;\n\tnoiseFloor: number | null;\n\tsampleCount: number;\n}\n\nexport interface ImprovementDecisionInput {\n\tcurrentMetric: number | null | undefined;\n\tbestMetric?: number | null;\n\tbaselineMetric?: number | null;\n\tdirection?: MetricDirection;\n\tchecksPass?: boolean | null;\n\tminDelta?: number;\n\tminRelativeDelta?: number;\n\tconfidenceMode?: ConfidenceMode;\n\tminConfidence?: number;\n\tlowConfidenceAction?: LowConfidenceAction;\n\tnoiseMetrics?: number[];\n}\n\nexport interface ImprovementDecisionResult {\n\tdecision: ImprovementDecision;\n\treason: ImprovementDecisionReason;\n\tcomparison: MetricComparison | null;\n\tconfidence: ConfidenceResult;\n}\n\nexport interface ImprovementRunRecord {\n\trunId: string | number;\n\tobjective?: string;\n\thypothesis?: string;\n\tmetricName: string;\n\tmetricUnit?: string;\n\tdirection: MetricDirection;\n\tmetric: number | null;\n\tsecondaryMetrics?: Record<string, number>;\n\tchecksPass: boolean | null;\n\tdecision: ImprovementDecision;\n\treason: ImprovementDecisionReason;\n\tconfidence?: number | null;\n\tchangedFiles?: string[];\n\tevidenceRef?: string;\n\tnextHint?: string;\n\ttimestamp: number;\n}\n\nexport interface GitStatusEntry {\n\tindex: string;\n\tworkingTree: string;\n\tpath: string;\n\torigPath?: string;\n}\n\nexport interface OwnedDiscardPlanInput {\n\tbeforeStatus: string | GitStatusEntry[];\n\tafterStatus: string | GitStatusEntry[];\n\townedPaths: string[];\n\tpreservePaths?: string[];\n}\n\nexport interface OwnedDiscardPlan {\n\trevertPaths: string[];\n\tpreservePaths: string[];\n\tprotectedUserDirtyPaths: string[];\n\tunownedChangedPaths: string[];\n\tcanDiscardOwnedChanges: boolean;\n}\n\nexport interface ImprovementLoopConfig {\n\tloopId: string;\n\tobjective: string;\n\tmetricName: string;\n\tmetricUnit?: string;\n\tdirection: MetricDirection;\n\tminDelta: number;\n\tminRelativeDelta: number;\n\tconfidenceMode: ConfidenceMode;\n\tminConfidence: number;\n\tlowConfidenceAction: LowConfidenceAction;\n\tcreatedAt: number;\n\tcwd: string;\n}\n\nexport interface ImprovementSandboxRecord {\n\tsandboxId: string;\n\tstatus: \"active\" | \"cleaned\";\n\trepoPath: string;\n\tworktreePath: string;\n\tbaseRef: string;\n\tcreatedAt: number;\n\tcleanedAt?: number;\n\texportedAt?: number;\n\tpatchPath?: string;\n\tpatchBytes?: number;\n\treason?: string;\n}\n\nexport interface ImprovementLoopState {\n\tconfig: ImprovementLoopConfig;\n\truns: ImprovementRunRecord[];\n\tsandboxes: ImprovementSandboxRecord[];\n\tactiveSandbox: ImprovementSandboxRecord | null;\n\tbaselineMetric: number | null;\n\tbestMetric: number | null;\n\tbestRunId: string | number | null;\n\tlastDecision: ImprovementDecisionResult | null;\n\tlogPath: string;\n}\n\nexport interface ImprovementLoopPaths {\n\trootDir: string;\n\tworkspaceDir: string;\n\tlogPath: string;\n\tsandboxDir: string;\n\tartifactDir: string;\n\tworkspaceKey: string;\n\tloopId: string;\n}\n\nexport interface ImprovementLoopInitInput {\n\tcwd: string;\n\tobjective: string;\n\tmetricName: string;\n\tmetricUnit?: string;\n\tdirection?: MetricDirection;\n\tloopId?: string;\n\tminDelta?: number;\n\tminRelativeDelta?: number;\n\tconfidenceMode?: ConfidenceMode;\n\tminConfidence?: number;\n\tlowConfidenceAction?: LowConfidenceAction;\n\tagentDir?: string;\n\treset?: boolean;\n}\n\nexport interface ImprovementLoopRecordInput {\n\tcwd: string;\n\tloopId?: string;\n\trunId?: string | number;\n\tmetric: number | null | undefined;\n\tsecondaryMetrics?: Record<string, number>;\n\tchecksPass?: boolean | null;\n\thypothesis?: string;\n\tchangedFiles?: string[];\n\tevidenceRef?: string;\n\tnextHint?: string;\n\tagentDir?: string;\n}\n\nexport interface ImprovementLoopStatusInput {\n\tcwd: string;\n\tloopId?: string;\n\tagentDir?: string;\n}\n\nexport interface ImprovementSandboxCreateInput extends ImprovementLoopStatusInput {\n\texec: ImprovementLoopExec;\n\tbaseRef?: string;\n\tallowDirtyRepo?: boolean;\n\tsandboxId?: string;\n\tsignal?: AbortSignal;\n}\n\nexport interface ImprovementSandboxCleanupInput extends ImprovementLoopStatusInput {\n\texec: ImprovementLoopExec;\n\tsandboxId?: string;\n\treason?: string;\n\tsignal?: AbortSignal;\n}\n\nexport interface ImprovementSandboxExportInput extends ImprovementLoopStatusInput {\n\texec: ImprovementLoopExec;\n\tsandboxId?: string;\n\tallowEmptyPatch?: boolean;\n\tsignal?: AbortSignal;\n}\n\ntype ImprovementLoopLogEntry =\n\t| ({ type: \"config\" } & ImprovementLoopConfig)\n\t| ({ type: \"run\" } & ImprovementRunRecord)\n\t| ({ type: \"sandbox\" } & ImprovementSandboxRecord);\n\nexport type ImprovementLoopToolAction =\n\t| \"init\"\n\t| \"status\"\n\t| \"record\"\n\t| \"measure\"\n\t| \"sandbox_create\"\n\t| \"sandbox_export\"\n\t| \"sandbox_cleanup\";\n\nexport interface ImprovementMeasurementCommandResult {\n\tstdout: string;\n\tstderr: string;\n\tcode: number;\n\tkilled: boolean;\n\tstdoutTruncated?: boolean;\n\tstderrTruncated?: boolean;\n}\n\nexport interface ImprovementMeasurementResult {\n\tcommand: string;\n\texitCode: number;\n\ttimedOut: boolean;\n\tdurationMs: number;\n\tstdout: string;\n\tstderr: string;\n\tstdoutTruncated: boolean;\n\tstderrTruncated: boolean;\n\tparsedMetrics: Record<string, number>;\n\tprimaryMetric: number | null;\n\tchecksCommand?: string;\n\tchecksExitCode?: number;\n\tchecksTimedOut?: boolean;\n\tchecksPass: boolean;\n\tchecksStdout?: string;\n\tchecksStderr?: string;\n}\n\nexport type ImprovementLoopExec = (\n\tcommand: string,\n\targs: string[],\n\toptions?: { cwd?: string; timeout?: number; signal?: AbortSignal; maxBuffer?: number },\n) => Promise<ImprovementMeasurementCommandResult>;\n\nexport interface ImprovementLoopToolDetails {\n\taction: ImprovementLoopToolAction;\n\tstate: ImprovementLoopState | null;\n\tdecision?: ImprovementDecisionResult;\n\tmeasurement?: ImprovementMeasurementResult;\n\tsandbox?: ImprovementSandboxRecord;\n\tlogPath: string;\n}\n\nconst METRIC_LINE_PREFIX = \"METRIC\";\nconst DENIED_METRIC_NAMES = new Set([\"__proto__\", \"constructor\", \"prototype\"]);\nconst METRIC_NAME_RE = /^[\\w.µ]+$/u;\nconst DECIMAL_NUMBER_RE = /^[+-]?(?:(?:\\d+(?:\\.\\d*)?)|(?:\\.\\d+))(?:[eE][+-]?\\d+)?$/;\n\nfunction normalizeFiniteNumber(value: unknown): number | null {\n\tif (typeof value !== \"number\" || !Number.isFinite(value)) return null;\n\treturn value;\n}\n\nfunction normalizeNonNegative(value: number | undefined, fallback: number): number {\n\tif (typeof value !== \"number\" || !Number.isFinite(value) || value < 0) return fallback;\n\treturn value;\n}\n\nfunction normalizeDirection(direction: MetricDirection | undefined): MetricDirection {\n\treturn direction === \"higher\" ? \"higher\" : \"lower\";\n}\n\nexport function parseMetricLines(output: string): Map<string, number> {\n\tconst metrics = new Map<string, number>();\n\tfor (const line of output.split(/\\r?\\n/)) {\n\t\tconst trimmed = line.trim();\n\t\tif (!trimmed.startsWith(`${METRIC_LINE_PREFIX} `)) continue;\n\t\tconst body = trimmed.slice(METRIC_LINE_PREFIX.length).trim();\n\t\tconst equals = body.indexOf(\"=\");\n\t\tif (equals <= 0) continue;\n\t\tconst name = body.slice(0, equals);\n\t\tconst rawValue = body.slice(equals + 1).trim();\n\t\tif (!METRIC_NAME_RE.test(name) || DENIED_METRIC_NAMES.has(name)) continue;\n\t\tif (!DECIMAL_NUMBER_RE.test(rawValue)) continue;\n\t\tconst value = Number(rawValue);\n\t\tif (Number.isFinite(value)) metrics.set(name, value);\n\t}\n\treturn metrics;\n}\n\nexport function metricMapFromOutput(output: string): Record<string, number> {\n\treturn Object.fromEntries(parseMetricLines(output));\n}\n\nexport function selectPrimaryMetric(\n\tmetrics: Map<string, number> | Record<string, number>,\n\tmetricName: string,\n): number | null {\n\tconst value = metrics instanceof Map ? metrics.get(metricName) : metrics[metricName];\n\treturn normalizeFiniteNumber(value);\n}\n\nexport function compareMetric(\n\tcurrent: number,\n\tbest: number | null | undefined,\n\tdirection: MetricDirection = \"lower\",\n): MetricComparison {\n\tconst normalizedBest = normalizeFiniteNumber(best) ?? null;\n\tif (normalizedBest === null) {\n\t\treturn {\n\t\t\tcurrent,\n\t\t\tbest: null,\n\t\t\tdirection,\n\t\t\tdelta: Number.POSITIVE_INFINITY,\n\t\t\trelativeDelta: null,\n\t\t\timproved: true,\n\t\t};\n\t}\n\n\tconst delta = direction === \"lower\" ? normalizedBest - current : current - normalizedBest;\n\tconst relativeDelta = normalizedBest === 0 ? null : delta / Math.abs(normalizedBest);\n\treturn {\n\t\tcurrent,\n\t\tbest: normalizedBest,\n\t\tdirection,\n\t\tdelta,\n\t\trelativeDelta,\n\t\timproved: delta > 0,\n\t};\n}\n\nexport function median(values: number[]): number | null {\n\tconst finite = values.filter(Number.isFinite).sort((a, b) => a - b);\n\tif (finite.length === 0) return null;\n\tconst mid = Math.floor(finite.length / 2);\n\treturn finite.length % 2 === 0 ? (finite[mid - 1] + finite[mid]) / 2 : finite[mid];\n}\n\nexport function medianAbsoluteDeviation(values: number[]): number | null {\n\tconst m = median(values);\n\tif (m === null) return null;\n\treturn median(values.filter(Number.isFinite).map((value) => Math.abs(value - m)));\n}\n\nexport function computeMadConfidence(improvementDelta: number, samples: number[]): ConfidenceResult {\n\tconst finite = samples.filter(Number.isFinite);\n\tif (finite.length < 3) {\n\t\treturn { mode: \"mad\", value: null, noiseFloor: null, sampleCount: finite.length };\n\t}\n\tconst mad = medianAbsoluteDeviation(finite);\n\tif (mad === null || mad === 0) {\n\t\treturn { mode: \"mad\", value: null, noiseFloor: mad, sampleCount: finite.length };\n\t}\n\treturn {\n\t\tmode: \"mad\",\n\t\tvalue: Math.abs(improvementDelta) / mad,\n\t\tnoiseFloor: mad,\n\t\tsampleCount: finite.length,\n\t};\n}\n\nexport function decideImprovement(input: ImprovementDecisionInput): ImprovementDecisionResult {\n\tconst direction = normalizeDirection(input.direction);\n\tconst minDelta = normalizeNonNegative(input.minDelta, 0);\n\tconst minRelativeDelta = normalizeNonNegative(input.minRelativeDelta, 0);\n\tconst minConfidence = normalizeNonNegative(input.minConfidence, 0);\n\tconst confidenceMode = input.confidenceMode ?? (minConfidence > 0 ? \"mad\" : \"none\");\n\tconst lowConfidenceAction = input.lowConfidenceAction ?? \"retry\";\n\n\tconst emptyConfidence: ConfidenceResult = { mode: confidenceMode, value: null, noiseFloor: null, sampleCount: 0 };\n\n\tif (input.checksPass === false) {\n\t\treturn { decision: \"discard\", reason: \"checks_failed\", comparison: null, confidence: emptyConfidence };\n\t}\n\n\tconst current = normalizeFiniteNumber(input.currentMetric);\n\tif (current === null) {\n\t\treturn {\n\t\t\tdecision: \"blocked\",\n\t\t\treason: input.currentMetric == null ? \"metric_missing\" : \"metric_invalid\",\n\t\t\tcomparison: null,\n\t\t\tconfidence: emptyConfidence,\n\t\t};\n\t}\n\n\tconst best = normalizeFiniteNumber(input.bestMetric) ?? null;\n\tconst comparison = compareMetric(current, best, direction);\n\tif (best === null) {\n\t\treturn { decision: \"keep\", reason: \"baseline\", comparison, confidence: emptyConfidence };\n\t}\n\tif (!comparison.improved) {\n\t\treturn { decision: \"discard\", reason: \"not_better_than_best\", comparison, confidence: emptyConfidence };\n\t}\n\tif (comparison.delta < minDelta) {\n\t\treturn { decision: \"discard\", reason: \"below_min_delta\", comparison, confidence: emptyConfidence };\n\t}\n\tif (comparison.relativeDelta !== null && comparison.relativeDelta < minRelativeDelta) {\n\t\treturn { decision: \"discard\", reason: \"below_min_relative_delta\", comparison, confidence: emptyConfidence };\n\t}\n\n\tif (confidenceMode === \"mad\" && minConfidence > 0) {\n\t\tconst samples = [...(input.noiseMetrics ?? []), current];\n\t\tconst confidence = computeMadConfidence(comparison.delta, samples);\n\t\tif (confidence.value === null) {\n\t\t\treturn {\n\t\t\t\tdecision: lowConfidenceAction,\n\t\t\t\treason: \"insufficient_noise_evidence\",\n\t\t\t\tcomparison,\n\t\t\t\tconfidence,\n\t\t\t};\n\t\t}\n\t\tif (confidence.value < minConfidence) {\n\t\t\treturn {\n\t\t\t\tdecision: lowConfidenceAction,\n\t\t\t\treason: \"below_confidence\",\n\t\t\t\tcomparison,\n\t\t\t\tconfidence,\n\t\t\t};\n\t\t}\n\t\treturn { decision: \"keep\", reason: \"metric_improved\", comparison, confidence };\n\t}\n\n\treturn { decision: \"keep\", reason: \"metric_improved\", comparison, confidence: emptyConfidence };\n}\n\nexport function serializeRunRecord(record: ImprovementRunRecord): string {\n\treturn JSON.stringify(record);\n}\n\nexport async function appendRunRecord(filePath: string, record: ImprovementRunRecord): Promise<void> {\n\tawait mkdir(dirname(filePath), { recursive: true });\n\tawait writeFile(filePath, `${serializeRunRecord(record)}\\n`, { flag: \"a\" });\n}\n\nexport async function readRunRecords(filePath: string): Promise<ImprovementRunRecord[]> {\n\tlet text = \"\";\n\ttry {\n\t\ttext = await readFile(filePath, \"utf8\");\n\t} catch (error) {\n\t\tif ((error as NodeJS.ErrnoException).code === \"ENOENT\") return [];\n\t\tthrow error;\n\t}\n\tconst records: ImprovementRunRecord[] = [];\n\tfor (const line of text.split(\"\\n\")) {\n\t\tif (!line.trim()) continue;\n\t\tconst parsed = JSON.parse(line) as ImprovementRunRecord;\n\t\trecords.push(parsed);\n\t}\n\treturn records;\n}\n\nfunction normalizeLoopId(loopId: string | undefined): string {\n\tconst normalized = (loopId ?? \"default\")\n\t\t.trim()\n\t\t.replace(/[^a-zA-Z0-9._-]+/g, \"-\")\n\t\t.replace(/^-+|-+$/g, \"\");\n\treturn normalized || \"default\";\n}\n\nfunction workspaceKeyFor(cwd: string): string {\n\treturn createHash(\"sha256\").update(cwd).digest(\"hex\").slice(0, 16);\n}\n\nexport function improvementLoopPaths(input: ImprovementLoopStatusInput): ImprovementLoopPaths {\n\tconst rootDir = join(input.agentDir ?? getAgentDir(), \"improvement-loop\");\n\tconst workspaceKey = workspaceKeyFor(input.cwd);\n\tconst loopId = normalizeLoopId(input.loopId);\n\tconst workspaceDir = join(rootDir, \"workspaces\", workspaceKey);\n\treturn {\n\t\trootDir,\n\t\tworkspaceDir,\n\t\tlogPath: join(workspaceDir, `${loopId}.jsonl`),\n\t\tsandboxDir: join(workspaceDir, \"sandboxes\", loopId),\n\t\tartifactDir: join(workspaceDir, \"artifacts\", loopId),\n\t\tworkspaceKey,\n\t\tloopId,\n\t};\n}\n\nfunction finiteRunMetric(run: ImprovementRunRecord): number | null {\n\treturn normalizeFiniteNumber(run.metric);\n}\n\nfunction betterRun(\n\tcurrent: ImprovementRunRecord,\n\tbest: ImprovementRunRecord | null,\n\tdirection: MetricDirection,\n): ImprovementRunRecord {\n\tif (!best) return current;\n\tconst currentMetric = finiteRunMetric(current);\n\tconst bestMetric = finiteRunMetric(best);\n\tif (currentMetric === null || bestMetric === null) return best;\n\treturn compareMetric(currentMetric, bestMetric, direction).improved ? current : best;\n}\n\nfunction configFromInit(input: ImprovementLoopInitInput, loopId: string): ImprovementLoopConfig {\n\treturn {\n\t\tloopId,\n\t\tobjective: input.objective,\n\t\tmetricName: input.metricName,\n\t\tmetricUnit: input.metricUnit ?? \"\",\n\t\tdirection: normalizeDirection(input.direction),\n\t\tminDelta: normalizeNonNegative(input.minDelta, 0),\n\t\tminRelativeDelta: normalizeNonNegative(input.minRelativeDelta, 0),\n\t\tconfidenceMode: input.confidenceMode ?? (normalizeNonNegative(input.minConfidence, 0) > 0 ? \"mad\" : \"none\"),\n\t\tminConfidence: normalizeNonNegative(input.minConfidence, 0),\n\t\tlowConfidenceAction: input.lowConfidenceAction ?? \"retry\",\n\t\tcreatedAt: Date.now(),\n\t\tcwd: input.cwd,\n\t};\n}\n\nasync function readLoopLogEntries(logPath: string): Promise<ImprovementLoopLogEntry[]> {\n\tlet text = \"\";\n\ttry {\n\t\ttext = await readFile(logPath, \"utf8\");\n\t} catch (error) {\n\t\tif ((error as NodeJS.ErrnoException).code === \"ENOENT\") return [];\n\t\tthrow error;\n\t}\n\treturn text\n\t\t.split(\"\\n\")\n\t\t.filter((line) => line.trim().length > 0)\n\t\t.map((line) => JSON.parse(line) as ImprovementLoopLogEntry);\n}\n\nfunction reconstructImprovementLoopState(\n\tentries: ImprovementLoopLogEntry[],\n\tlogPath: string,\n): ImprovementLoopState | null {\n\tlet config: ImprovementLoopConfig | null = null;\n\tlet runs: ImprovementRunRecord[] = [];\n\tconst sandboxById = new Map<string, ImprovementSandboxRecord>();\n\tfor (const entry of entries) {\n\t\tif (entry.type === \"config\") {\n\t\t\tconst { type: _type, ...rest } = entry;\n\t\t\tconfig = rest;\n\t\t\truns = [];\n\t\t\tsandboxById.clear();\n\t\t\tcontinue;\n\t\t}\n\t\tif (entry.type === \"run\") {\n\t\t\tconst { type: _type, ...run } = entry;\n\t\t\truns.push(run);\n\t\t\tcontinue;\n\t\t}\n\t\tif (entry.type === \"sandbox\") {\n\t\t\tconst { type: _type, ...sandbox } = entry;\n\t\t\tsandboxById.set(sandbox.sandboxId, sandbox);\n\t\t}\n\t}\n\tif (!config) return null;\n\n\tconst baselineMetric = finiteRunMetric(runs[0] ?? ({} as ImprovementRunRecord));\n\tlet bestRun: ImprovementRunRecord | null = null;\n\tfor (const run of runs) {\n\t\tif (run.decision !== \"keep\" || finiteRunMetric(run) === null) continue;\n\t\tbestRun = betterRun(run, bestRun, config.direction);\n\t}\n\n\tconst sandboxes = [...sandboxById.values()].sort((a, b) => a.createdAt - b.createdAt);\n\tconst activeSandbox = [...sandboxes].reverse().find((sandbox) => sandbox.status === \"active\") ?? null;\n\treturn {\n\t\tconfig,\n\t\truns,\n\t\tsandboxes,\n\t\tactiveSandbox,\n\t\tbaselineMetric,\n\t\tbestMetric: bestRun ? finiteRunMetric(bestRun) : null,\n\t\tbestRunId: bestRun?.runId ?? null,\n\t\tlastDecision: null,\n\t\tlogPath,\n\t};\n}\n\nexport async function initImprovementLoop(input: ImprovementLoopInitInput): Promise<ImprovementLoopState> {\n\tconst paths = improvementLoopPaths(input);\n\tconst existing = await readLoopLogEntries(paths.logPath);\n\tif (existing.length > 0 && !input.reset) {\n\t\tthrow new Error(\n\t\t\t`Improvement loop already exists at ${paths.logPath}; pass reset=true to replace user-level loop state.`,\n\t\t);\n\t}\n\tconst config = configFromInit(input, paths.loopId);\n\tawait mkdir(dirname(paths.logPath), { recursive: true });\n\tawait writeFile(paths.logPath, `${JSON.stringify({ type: \"config\", ...config })}\\n`);\n\tconst state = reconstructImprovementLoopState([{ type: \"config\", ...config }], paths.logPath);\n\tif (!state) throw new Error(\"Failed to initialize improvement loop state\");\n\treturn state;\n}\n\nexport async function readImprovementLoopState(\n\tinput: ImprovementLoopStatusInput,\n): Promise<ImprovementLoopState | null> {\n\tconst paths = improvementLoopPaths(input);\n\treturn reconstructImprovementLoopState(await readLoopLogEntries(paths.logPath), paths.logPath);\n}\n\nfunction createSandboxId(input: ImprovementSandboxCreateInput): string {\n\tif (input.sandboxId) return normalizeLoopId(input.sandboxId);\n\tconst seed = `${input.cwd}:${input.loopId ?? \"default\"}:${Date.now()}:${Math.random()}`;\n\treturn `${Date.now()}-${createHash(\"sha256\").update(seed).digest(\"hex\").slice(0, 8)}`;\n}\n\nasync function appendSandboxRecord(\n\tinput: ImprovementLoopStatusInput,\n\tsandbox: ImprovementSandboxRecord,\n): Promise<ImprovementLoopState> {\n\tconst paths = improvementLoopPaths(input);\n\tawait mkdir(dirname(paths.logPath), { recursive: true });\n\tawait writeFile(paths.logPath, `${JSON.stringify({ type: \"sandbox\", ...sandbox })}\\n`, { flag: \"a\" });\n\tconst updated = await readImprovementLoopState(input);\n\tif (!updated) throw new Error(\"Failed to read improvement loop after sandbox update\");\n\treturn updated;\n}\n\nasync function runGit(\n\texec: ImprovementLoopExec,\n\tcwd: string,\n\targs: string[],\n\tsignal?: AbortSignal,\n): Promise<ImprovementMeasurementCommandResult> {\n\treturn exec(\"git\", args, { cwd, timeout: 60_000, signal, maxBuffer: 64 * 1024 });\n}\n\nexport async function createImprovementSandbox(input: ImprovementSandboxCreateInput): Promise<ImprovementLoopState> {\n\tconst state = await readImprovementLoopState(input);\n\tif (!state) throw new Error(\"Improvement loop is not initialized; call init first.\");\n\tif (state.activeSandbox) {\n\t\tthrow new Error(\n\t\t\t`Active sandbox already exists at ${state.activeSandbox.worktreePath}; clean it before creating another.`,\n\t\t);\n\t}\n\tconst status = await runGit(input.exec, input.cwd, [\"status\", \"--porcelain\"], input.signal);\n\tif (status.code !== 0) throw new Error(`Cannot inspect git status: ${(status.stderr || status.stdout).trim()}`);\n\tif (status.stdout.trim() && !input.allowDirtyRepo) {\n\t\tthrow new Error(\n\t\t\t\"Refusing to create sandbox from dirty repository; commit/stash changes or pass allowDirtyRepo=true.\",\n\t\t);\n\t}\n\tconst baseRef = input.baseRef?.trim() || \"HEAD\";\n\tconst sandboxId = createSandboxId(input);\n\tconst paths = improvementLoopPaths(input);\n\tconst worktreePath = join(paths.sandboxDir, sandboxId);\n\tawait mkdir(dirname(worktreePath), { recursive: true });\n\tconst add = await runGit(\n\t\tinput.exec,\n\t\tinput.cwd,\n\t\t[\"worktree\", \"add\", \"--detach\", worktreePath, baseRef],\n\t\tinput.signal,\n\t);\n\tif (add.code !== 0) throw new Error(`Failed to create git worktree sandbox: ${(add.stderr || add.stdout).trim()}`);\n\tconst sandbox: ImprovementSandboxRecord = {\n\t\tsandboxId,\n\t\tstatus: \"active\",\n\t\trepoPath: input.cwd,\n\t\tworktreePath,\n\t\tbaseRef,\n\t\tcreatedAt: Date.now(),\n\t};\n\treturn appendSandboxRecord(input, sandbox);\n}\n\nexport async function exportImprovementSandboxPatch(\n\tinput: ImprovementSandboxExportInput,\n): Promise<ImprovementLoopState> {\n\tconst state = await readImprovementLoopState(input);\n\tif (!state) throw new Error(\"Improvement loop is not initialized; call init first.\");\n\tconst sandbox = input.sandboxId\n\t\t? state.sandboxes.find((candidate) => candidate.sandboxId === normalizeLoopId(input.sandboxId))\n\t\t: state.activeSandbox;\n\tif (!sandbox || sandbox.status !== \"active\") throw new Error(\"No active sandbox found to export.\");\n\tconst diff = await runGit(input.exec, sandbox.worktreePath, [\"diff\", \"--binary\", \"HEAD\"], input.signal);\n\tif (diff.code !== 0) throw new Error(`Failed to export sandbox patch: ${(diff.stderr || diff.stdout).trim()}`);\n\tif (!diff.stdout.trim() && !input.allowEmptyPatch) throw new Error(\"Sandbox has no changes to export.\");\n\tconst paths = improvementLoopPaths(input);\n\tawait mkdir(paths.artifactDir, { recursive: true });\n\tconst patchPath = join(paths.artifactDir, `${sandbox.sandboxId}.patch`);\n\tawait writeFile(patchPath, diff.stdout);\n\treturn appendSandboxRecord(input, {\n\t\t...sandbox,\n\t\texportedAt: Date.now(),\n\t\tpatchPath,\n\t\tpatchBytes: Buffer.byteLength(diff.stdout),\n\t});\n}\n\nexport async function cleanupImprovementSandbox(input: ImprovementSandboxCleanupInput): Promise<ImprovementLoopState> {\n\tconst state = await readImprovementLoopState(input);\n\tif (!state) throw new Error(\"Improvement loop is not initialized; call init first.\");\n\tconst sandbox = input.sandboxId\n\t\t? state.sandboxes.find((candidate) => candidate.sandboxId === normalizeLoopId(input.sandboxId))\n\t\t: state.activeSandbox;\n\tif (!sandbox || sandbox.status !== \"active\") throw new Error(\"No active sandbox found to clean up.\");\n\tconst remove = await runGit(\n\t\tinput.exec,\n\t\tsandbox.repoPath,\n\t\t[\"worktree\", \"remove\", \"--force\", sandbox.worktreePath],\n\t\tinput.signal,\n\t);\n\tif (remove.code !== 0)\n\t\tthrow new Error(`Failed to remove git worktree sandbox: ${(remove.stderr || remove.stdout).trim()}`);\n\tawait runGit(input.exec, sandbox.repoPath, [\"worktree\", \"prune\"], input.signal);\n\treturn appendSandboxRecord(input, {\n\t\t...sandbox,\n\t\tstatus: \"cleaned\",\n\t\tcleanedAt: Date.now(),\n\t\treason: input.reason,\n\t});\n}\n\nexport async function recordImprovementRun(input: ImprovementLoopRecordInput): Promise<ImprovementLoopState> {\n\tconst state = await readImprovementLoopState(input);\n\tif (!state) throw new Error(\"Improvement loop is not initialized; call init first.\");\n\tconst runId = input.runId ?? state.runs.length + 1;\n\tconst priorMetrics = state.runs\n\t\t.map((run) => run.metric)\n\t\t.filter((metric): metric is number => Number.isFinite(metric));\n\tconst decision = decideImprovement({\n\t\tcurrentMetric: input.metric,\n\t\tbestMetric: state.bestMetric,\n\t\tdirection: state.config.direction,\n\t\tchecksPass: input.checksPass ?? null,\n\t\tminDelta: state.config.minDelta,\n\t\tminRelativeDelta: state.config.minRelativeDelta,\n\t\tconfidenceMode: state.config.confidenceMode,\n\t\tminConfidence: state.config.minConfidence,\n\t\tlowConfidenceAction: state.config.lowConfidenceAction,\n\t\tnoiseMetrics: priorMetrics,\n\t});\n\tconst record: ImprovementRunRecord = {\n\t\trunId,\n\t\tobjective: state.config.objective,\n\t\thypothesis: input.hypothesis,\n\t\tmetricName: state.config.metricName,\n\t\tmetricUnit: state.config.metricUnit,\n\t\tdirection: state.config.direction,\n\t\tmetric: normalizeFiniteNumber(input.metric),\n\t\tsecondaryMetrics: input.secondaryMetrics,\n\t\tchecksPass: input.checksPass ?? null,\n\t\tdecision: decision.decision,\n\t\treason: decision.reason,\n\t\tconfidence: decision.confidence.value,\n\t\tchangedFiles: input.changedFiles,\n\t\tevidenceRef: input.evidenceRef,\n\t\tnextHint: input.nextHint,\n\t\ttimestamp: Date.now(),\n\t};\n\tawait mkdir(dirname(state.logPath), { recursive: true });\n\tawait writeFile(state.logPath, `${JSON.stringify({ type: \"run\", ...record })}\\n`, { flag: \"a\" });\n\tconst updated = await readImprovementLoopState(input);\n\tif (!updated) throw new Error(\"Failed to read improvement loop after recording run\");\n\treturn { ...updated, lastDecision: decision };\n}\n\nfunction secondaryMetricsFrom(\n\tparsedMetrics: Record<string, number>,\n\tprimaryName: string,\n): Record<string, number> | undefined {\n\tconst secondary = Object.fromEntries(Object.entries(parsedMetrics).filter(([name]) => name !== primaryName));\n\treturn Object.keys(secondary).length > 0 ? secondary : undefined;\n}\n\nexport async function runImprovementMeasurement(input: {\n\texec: ImprovementLoopExec;\n\tcwd: string;\n\tcommand: string;\n\tmetricName: string;\n\tchecksCommand?: string;\n\ttimeoutSeconds?: number;\n\tchecksTimeoutSeconds?: number;\n\tmaxOutputBytes?: number;\n\tsignal?: AbortSignal;\n}): Promise<ImprovementMeasurementResult> {\n\tconst command = input.command.trim();\n\tif (!command) throw new Error(\"command is required\");\n\tconst timeoutMs = Math.max(1, Math.floor(input.timeoutSeconds ?? 600)) * 1000;\n\tconst checksTimeoutMs = Math.max(1, Math.floor(input.checksTimeoutSeconds ?? 300)) * 1000;\n\tconst maxBuffer = Math.max(1024, Math.floor(input.maxOutputBytes ?? 64 * 1024));\n\tconst started = Date.now();\n\tconst result = await input.exec(\"bash\", [\"-c\", command], {\n\t\tcwd: input.cwd,\n\t\ttimeout: timeoutMs,\n\t\tsignal: input.signal,\n\t\tmaxBuffer,\n\t});\n\tconst durationMs = Date.now() - started;\n\tconst output = `${result.stdout}\\n${result.stderr}`;\n\tconst parsedMetrics = metricMapFromOutput(output);\n\tconst primaryMetric = selectPrimaryMetric(parsedMetrics, input.metricName);\n\tlet checksPass = result.code === 0 && !result.killed;\n\tlet checksExitCode: number | undefined;\n\tlet checksTimedOut: boolean | undefined;\n\tlet checksStdout: string | undefined;\n\tlet checksStderr: string | undefined;\n\tconst checksCommand = input.checksCommand?.trim();\n\tif (checksPass && checksCommand) {\n\t\tconst checks = await input.exec(\"bash\", [\"-c\", checksCommand], {\n\t\t\tcwd: input.cwd,\n\t\t\ttimeout: checksTimeoutMs,\n\t\t\tsignal: input.signal,\n\t\t\tmaxBuffer,\n\t\t});\n\t\tchecksExitCode = checks.code;\n\t\tchecksTimedOut = checks.killed;\n\t\tchecksStdout = checks.stdout;\n\t\tchecksStderr = checks.stderr;\n\t\tchecksPass = checks.code === 0 && !checks.killed;\n\t}\n\n\treturn {\n\t\tcommand,\n\t\texitCode: result.code,\n\t\ttimedOut: result.killed,\n\t\tdurationMs,\n\t\tstdout: result.stdout,\n\t\tstderr: result.stderr,\n\t\tstdoutTruncated: !!result.stdoutTruncated,\n\t\tstderrTruncated: !!result.stderrTruncated,\n\t\tparsedMetrics,\n\t\tprimaryMetric,\n\t\tchecksCommand: checksCommand || undefined,\n\t\tchecksExitCode,\n\t\tchecksTimedOut,\n\t\tchecksPass,\n\t\tchecksStdout,\n\t\tchecksStderr,\n\t};\n}\n\nexport function parseGitPorcelainStatus(status: string): GitStatusEntry[] {\n\treturn status\n\t\t.split(\"\\n\")\n\t\t.map((line) => line.trimEnd())\n\t\t.filter(Boolean)\n\t\t.map((line) => {\n\t\t\tconst index = line[0] ?? \" \";\n\t\t\tconst workingTree = line[1] ?? \" \";\n\t\t\tconst rawPath = line.slice(3);\n\t\t\tconst renameParts = rawPath.split(\" -> \");\n\t\t\tif (renameParts.length === 2) {\n\t\t\t\treturn { index, workingTree, origPath: renameParts[0], path: renameParts[1] };\n\t\t\t}\n\t\t\treturn { index, workingTree, path: rawPath };\n\t\t});\n}\n\nfunction entriesFromStatus(status: string | GitStatusEntry[]): GitStatusEntry[] {\n\treturn Array.isArray(status) ? status : parseGitPorcelainStatus(status);\n}\n\nfunction normalizePath(path: string): string {\n\treturn path.replace(/\\\\/g, \"/\").replace(/^\\.\\//, \"\");\n}\n\nfunction pathSet(paths: string[]): Set<string> {\n\treturn new Set(paths.map(normalizePath).filter(Boolean));\n}\n\nfunction entryPaths(entry: GitStatusEntry): string[] {\n\treturn [entry.path, entry.origPath].filter((value): value is string => typeof value === \"string\").map(normalizePath);\n}\n\nfunction statusPathSet(entries: GitStatusEntry[]): Set<string> {\n\tconst paths = new Set<string>();\n\tfor (const entry of entries) {\n\t\tfor (const path of entryPaths(entry)) paths.add(path);\n\t}\n\treturn paths;\n}\n\nfunction isPreserved(path: string, preservePaths: Set<string>): boolean {\n\tfor (const preserve of preservePaths) {\n\t\tif (path === preserve || path.startsWith(`${preserve}/`)) return true;\n\t}\n\treturn false;\n}\n\nexport function planOwnedDiscard(input: OwnedDiscardPlanInput): OwnedDiscardPlan {\n\tconst beforePaths = statusPathSet(entriesFromStatus(input.beforeStatus));\n\tconst afterPaths = statusPathSet(entriesFromStatus(input.afterStatus));\n\tconst ownedPaths = pathSet(input.ownedPaths);\n\tconst preservePaths = pathSet(input.preservePaths ?? []);\n\tconst revertPaths: string[] = [];\n\tconst protectedUserDirtyPaths: string[] = [];\n\tconst unownedChangedPaths: string[] = [];\n\tconst preservedChangedPaths: string[] = [];\n\n\tfor (const path of afterPaths) {\n\t\tif (isPreserved(path, preservePaths)) {\n\t\t\tpreservedChangedPaths.push(path);\n\t\t\tcontinue;\n\t\t}\n\t\tif (!ownedPaths.has(path)) {\n\t\t\tunownedChangedPaths.push(path);\n\t\t\tcontinue;\n\t\t}\n\t\tif (beforePaths.has(path)) {\n\t\t\tprotectedUserDirtyPaths.push(path);\n\t\t\tcontinue;\n\t\t}\n\t\trevertPaths.push(path);\n\t}\n\n\treturn {\n\t\trevertPaths: [...new Set(revertPaths)].sort(),\n\t\tpreservePaths: [...new Set(preservedChangedPaths)].sort(),\n\t\tprotectedUserDirtyPaths: [...new Set(protectedUserDirtyPaths)].sort(),\n\t\tunownedChangedPaths: [...new Set(unownedChangedPaths)].sort(),\n\t\tcanDiscardOwnedChanges: protectedUserDirtyPaths.length === 0,\n\t};\n}\n\nconst DecisionToolParams = Type.Object(\n\t{\n\t\tcurrentMetric: Type.Number({ description: \"Measured primary metric for the candidate run.\" }),\n\t\tbestMetric: Type.Optional(Type.Number({ description: \"Best kept primary metric so far. Omit for baseline.\" })),\n\t\tdirection: StringEnum([\"lower\", \"higher\"] as const, {\n\t\t\tdescription: \"Whether lower or higher metric values are better. Default lower.\",\n\t\t}),\n\t\tchecksPass: Type.Optional(\n\t\t\tType.Boolean({ description: \"Whether correctness checks passed. false forces discard.\" }),\n\t\t),\n\t\tminDelta: Type.Optional(\n\t\t\tType.Number({ description: \"Minimum absolute improvement required to keep. Default 0.\" }),\n\t\t),\n\t\tminRelativeDelta: Type.Optional(\n\t\t\tType.Number({ description: \"Minimum relative improvement required to keep, e.g. 0.01 for 1%. Default 0.\" }),\n\t\t),\n\t\tconfidenceMode: Type.Optional(\n\t\t\tStringEnum([\"none\", \"mad\"] as const, {\n\t\t\t\tdescription: \"Noise/confidence policy. Use mad with minConfidence to reject noisy wins.\",\n\t\t\t}),\n\t\t),\n\t\tminConfidence: Type.Optional(\n\t\t\tType.Number({ description: \"Minimum confidence multiple over MAD noise floor. Default 0.\" }),\n\t\t),\n\t\tlowConfidenceAction: Type.Optional(\n\t\t\tStringEnum([\"retry\", \"discard\"] as const, {\n\t\t\t\tdescription: \"Decision when confidence evidence is missing or too low. Default retry.\",\n\t\t\t}),\n\t\t),\n\t\tnoiseMetrics: Type.Optional(\n\t\t\tType.Array(Type.Number(), { description: \"Previous metric samples for MAD confidence calculation.\" }),\n\t\t),\n\t},\n\t{ additionalProperties: false },\n);\n\nexport function createImprovementDecisionTool(): ToolDefinition<typeof DecisionToolParams, ImprovementDecisionResult> {\n\treturn defineTool({\n\t\tname: \"improvement_decision\",\n\t\tlabel: \"Improvement Decision\",\n\t\tdescription:\n\t\t\t\"Deterministically decide keep/discard/retry/blocked for an improvement candidate from metric, direction, thresholds, checks, and optional noise samples. The model proposes; code judges.\",\n\t\tpromptSnippet: \"Decide keep/discard for a measured improvement candidate using deterministic metric gates\",\n\t\tpromptGuidelines: [\n\t\t\t\"Use improvement_decision after measuring a candidate when keep/discard must be based on metric evidence rather than model intuition.\",\n\t\t\t\"Do not claim a candidate is kept unless this tool returns decision=keep or a stronger project-specific validator passes.\",\n\t\t\t\"Correctness failures override metric wins; checksPass=false always discards.\",\n\t\t],\n\t\tparameters: DecisionToolParams,\n\t\tasync execute(_toolCallId, params) {\n\t\t\tconst decision = decideImprovement(params);\n\t\t\tconst comparison = decision.comparison;\n\t\t\tconst confidenceText = decision.confidence.value === null ? \"n/a\" : decision.confidence.value.toFixed(2);\n\t\t\tconst deltaText = comparison ? comparison.delta.toString() : \"n/a\";\n\t\t\treturn {\n\t\t\t\tcontent: [\n\t\t\t\t\t{\n\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\ttext: `Decision: ${decision.decision} (${decision.reason})\\nDelta: ${deltaText}\\nConfidence: ${confidenceText}`,\n\t\t\t\t\t},\n\t\t\t\t],\n\t\t\t\tdetails: decision,\n\t\t\t};\n\t\t},\n\t});\n}\n\nconst LoopToolParams = Type.Object(\n\t{\n\t\taction: StringEnum(\n\t\t\t[\"init\", \"status\", \"record\", \"measure\", \"sandbox_create\", \"sandbox_export\", \"sandbox_cleanup\"] as const,\n\t\t\t{\n\t\t\t\tdescription:\n\t\t\t\t\t\"init creates user-level loop state, status reads it, record appends a supplied measurement, measure runs and records a bounded command, sandbox_create/export/cleanup manage disposable git worktrees and keep patches.\",\n\t\t\t},\n\t\t),\n\t\tloopId: Type.Optional(Type.String({ description: \"Loop id within the current workspace. Default: default.\" })),\n\t\tobjective: Type.Optional(Type.String({ description: \"Loop objective. Required for action=init.\" })),\n\t\tmetricName: Type.Optional(Type.String({ description: \"Primary metric name. Required for action=init.\" })),\n\t\tmetricUnit: Type.Optional(Type.String({ description: \"Display unit for the primary metric.\" })),\n\t\tdirection: Type.Optional(\n\t\t\tStringEnum([\"lower\", \"higher\"] as const, {\n\t\t\t\tdescription: \"Whether lower or higher metric values are better. Default lower.\",\n\t\t\t}),\n\t\t),\n\t\treset: Type.Optional(\n\t\t\tType.Boolean({\n\t\t\t\tdescription: \"For action=init, replace existing user-level loop state for this workspace/loop id.\",\n\t\t\t}),\n\t\t),\n\t\tminDelta: Type.Optional(\n\t\t\tType.Number({ description: \"Minimum absolute improvement required to keep. Default 0.\" }),\n\t\t),\n\t\tminRelativeDelta: Type.Optional(\n\t\t\tType.Number({ description: \"Minimum relative improvement required to keep. Default 0.\" }),\n\t\t),\n\t\tconfidenceMode: Type.Optional(\n\t\t\tStringEnum([\"none\", \"mad\"] as const, {\n\t\t\t\tdescription: \"Noise/confidence policy. Default none unless minConfidence > 0.\",\n\t\t\t}),\n\t\t),\n\t\tminConfidence: Type.Optional(\n\t\t\tType.Number({ description: \"Minimum confidence multiple over MAD noise floor. Default 0.\" }),\n\t\t),\n\t\tlowConfidenceAction: Type.Optional(\n\t\t\tStringEnum([\"retry\", \"discard\"] as const, {\n\t\t\t\tdescription: \"Decision when confidence evidence is missing or low. Default retry.\",\n\t\t\t}),\n\t\t),\n\t\tcurrentMetric: Type.Optional(Type.Number({ description: \"Measured primary metric for action=record.\" })),\n\t\tchecksPass: Type.Optional(\n\t\t\tType.Boolean({ description: \"Correctness check result for action=record. false forces discard.\" }),\n\t\t),\n\t\thypothesis: Type.Optional(Type.String({ description: \"What this run tried.\" })),\n\t\tsecondaryMetrics: Type.Optional(\n\t\t\tType.Object(\n\t\t\t\t{},\n\t\t\t\t{ additionalProperties: Type.Number(), description: \"Secondary metric map for action=record.\" },\n\t\t\t),\n\t\t),\n\t\tchangedFiles: Type.Optional(Type.Array(Type.String(), { description: \"Files changed by this candidate run.\" })),\n\t\tevidenceRef: Type.Optional(\n\t\t\tType.String({ description: \"Path, command, or artifact reference proving the measurement/check result.\" }),\n\t\t),\n\t\tnextHint: Type.Optional(\n\t\t\tType.String({ description: \"Useful next-step hint for later runs, especially after discard/retry/blocked.\" }),\n\t\t),\n\t\tcommand: Type.Optional(\n\t\t\tType.String({\n\t\t\t\tdescription:\n\t\t\t\t\t\"Measurement command for action=measure. Runs via bash -c in ctx.cwd or active sandbox when useSandbox=true.\",\n\t\t\t}),\n\t\t),\n\t\tchecksCommand: Type.Optional(\n\t\t\tType.String({\n\t\t\t\tdescription:\n\t\t\t\t\t\"Optional correctness command for action=measure. Runs only if the measurement command exits 0.\",\n\t\t\t}),\n\t\t),\n\t\ttimeoutSeconds: Type.Optional(\n\t\t\tType.Number({ description: \"Measurement command timeout in seconds. Default 600.\" }),\n\t\t),\n\t\tchecksTimeoutSeconds: Type.Optional(\n\t\t\tType.Number({ description: \"Checks command timeout in seconds. Default 300.\" }),\n\t\t),\n\t\tmaxOutputBytes: Type.Optional(\n\t\t\tType.Number({ description: \"Maximum stdout/stderr tail retained per command. Default 65536.\" }),\n\t\t),\n\t\tuseSandbox: Type.Optional(\n\t\t\tType.Boolean({\n\t\t\t\tdescription: \"For action=measure, run command/checks inside the active sandbox worktree. Default false.\",\n\t\t\t}),\n\t\t),\n\t\tsandboxId: Type.Optional(\n\t\t\tType.String({\n\t\t\t\tdescription:\n\t\t\t\t\t\"Sandbox id for sandbox_create/sandbox_export/sandbox_cleanup. Defaults to generated id or active sandbox.\",\n\t\t\t}),\n\t\t),\n\t\tbaseRef: Type.Optional(Type.String({ description: \"Git base ref for sandbox_create. Default HEAD.\" })),\n\t\tallowDirtyRepo: Type.Optional(\n\t\t\tType.Boolean({\n\t\t\t\tdescription: \"Allow sandbox_create when the real repo has uncommitted changes. Default false.\",\n\t\t\t}),\n\t\t),\n\t\tcleanupReason: Type.Optional(\n\t\t\tType.String({ description: \"Optional reason recorded when sandbox_cleanup removes a sandbox.\" }),\n\t\t),\n\t\tallowEmptyPatch: Type.Optional(\n\t\t\tType.Boolean({ description: \"For sandbox_export, allow writing an empty patch. Default false.\" }),\n\t\t),\n\t},\n\t{ additionalProperties: false },\n);\n\nfunction requireString(value: string | undefined, label: string): string {\n\tconst trimmed = value?.trim();\n\tif (!trimmed) throw new Error(`${label} is required`);\n\treturn trimmed;\n}\n\nfunction summarizeLoopState(state: ImprovementLoopState | null): string {\n\tif (!state) return \"Improvement loop is not initialized.\";\n\tconst lines = [\n\t\t`Loop ${state.config.loopId}: ${state.config.objective}`,\n\t\t`Metric: ${state.config.metricName} (${state.config.metricUnit || \"unitless\"}, ${state.config.direction} is better)`,\n\t\t`Runs: ${state.runs.length}`,\n\t\t`Baseline: ${state.baselineMetric ?? \"none\"}`,\n\t\t`Best: ${state.bestMetric ?? \"none\"}${state.bestRunId === null ? \"\" : ` (#${state.bestRunId})`}`,\n\t\t`Log: ${state.logPath}`,\n\t];\n\tconst lastRun = state.runs.at(-1);\n\tif (lastRun) lines.push(`Last: ${lastRun.decision} (${lastRun.reason}) metric=${lastRun.metric ?? \"missing\"}`);\n\tif (state.activeSandbox) lines.push(`Active sandbox: ${state.activeSandbox.worktreePath}`);\n\treturn lines.join(\"\\n\");\n}\n\nexport function createImprovementLoopTool(\n\texec?: ImprovementLoopExec,\n): ToolDefinition<typeof LoopToolParams, ImprovementLoopToolDetails> {\n\treturn defineTool({\n\t\tname: \"improvement_loop\",\n\t\tlabel: \"Improvement Loop\",\n\t\tdescription:\n\t\t\t\"Persist and evaluate a deterministic improvement loop in user-level state: init objective/metric policy, measure bounded commands, record measured runs, manage disposable git worktree sandboxes, and status current baseline/best/decision log. Does not commit/apply/revert real-repo files.\",\n\t\tpromptSnippet: \"Track deterministic improvement-loop state and record measured keep/discard decisions\",\n\t\tpromptGuidelines: [\n\t\t\t\"Use improvement_loop init before iterative optimization that needs metric-based keep/discard state.\",\n\t\t\t\"Use improvement_loop measure to run a bounded measurement command and optional checks command, then record the deterministic decision.\",\n\t\t\t\"Use improvement_loop sandbox_create before risky self-modifying experiments so edits happen in a disposable git worktree, not the real repository.\",\n\t\t\t\"Use improvement_loop sandbox_export to capture the sandbox diff as a user-level patch artifact before cleanup when the decision is keep.\",\n\t\t\t\"Use improvement_loop sandbox_cleanup after discard or after exporting/approving a patch; cleanup removes the disposable worktree record, not the real repository.\",\n\t\t\t\"Use improvement_loop record when measurement already happened elsewhere; code decides keep/discard/retry/blocked from metrics and gates.\",\n\t\t\t\"This tool does not commit/apply/revert real-repo files; keep/discard is a decision record until a later approved executor applies it.\",\n\t\t\t\"Operational state is stored under the user-level agent directory, not in the target repository.\",\n\t\t],\n\t\tparameters: LoopToolParams,\n\t\texecutionMode: \"sequential\",\n\t\tasync execute(_toolCallId, params, signal, _onUpdate, ctx) {\n\t\t\tconst cwd = ctx.cwd;\n\t\t\tif (params.action === \"init\") {\n\t\t\t\tconst state = await initImprovementLoop({\n\t\t\t\t\tcwd,\n\t\t\t\t\tloopId: params.loopId,\n\t\t\t\t\tobjective: requireString(params.objective, \"objective\"),\n\t\t\t\t\tmetricName: requireString(params.metricName, \"metricName\"),\n\t\t\t\t\tmetricUnit: params.metricUnit,\n\t\t\t\t\tdirection: params.direction,\n\t\t\t\t\tminDelta: params.minDelta,\n\t\t\t\t\tminRelativeDelta: params.minRelativeDelta,\n\t\t\t\t\tconfidenceMode: params.confidenceMode,\n\t\t\t\t\tminConfidence: params.minConfidence,\n\t\t\t\t\tlowConfidenceAction: params.lowConfidenceAction,\n\t\t\t\t\treset: params.reset,\n\t\t\t\t});\n\t\t\t\treturn {\n\t\t\t\t\tcontent: [{ type: \"text\", text: `Initialized improvement loop.\\n${summarizeLoopState(state)}` }],\n\t\t\t\t\tdetails: { action: \"init\", state, logPath: state.logPath },\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tif (params.action === \"record\") {\n\t\t\t\tconst state = await recordImprovementRun({\n\t\t\t\t\tcwd,\n\t\t\t\t\tloopId: params.loopId,\n\t\t\t\t\tmetric: params.currentMetric,\n\t\t\t\t\tsecondaryMetrics: params.secondaryMetrics as Record<string, number> | undefined,\n\t\t\t\t\tchecksPass: params.checksPass ?? null,\n\t\t\t\t\thypothesis: params.hypothesis,\n\t\t\t\t\tchangedFiles: params.changedFiles,\n\t\t\t\t\tevidenceRef: params.evidenceRef,\n\t\t\t\t\tnextHint: params.nextHint,\n\t\t\t\t});\n\t\t\t\tconst decision = state.lastDecision;\n\t\t\t\treturn {\n\t\t\t\t\tcontent: [\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: `Recorded run: ${decision?.decision ?? \"unknown\"} (${decision?.reason ?? \"unknown\"}).\\n${summarizeLoopState(state)}`,\n\t\t\t\t\t\t},\n\t\t\t\t\t],\n\t\t\t\t\tdetails: { action: \"record\", state, decision: decision ?? undefined, logPath: state.logPath },\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tif (params.action === \"measure\") {\n\t\t\t\tif (!exec) throw new Error(\"improvement_loop measure requires createImprovementLoopTool(pi.exec)\");\n\t\t\t\tconst before = await readImprovementLoopState({ cwd, loopId: params.loopId });\n\t\t\t\tif (!before) throw new Error(\"Improvement loop is not initialized; call init first.\");\n\t\t\t\tconst measurementCwd = params.useSandbox\n\t\t\t\t\t? (before.activeSandbox?.worktreePath ??\n\t\t\t\t\t\t(() => {\n\t\t\t\t\t\t\tthrow new Error(\"useSandbox=true requires an active sandbox.\");\n\t\t\t\t\t\t})())\n\t\t\t\t\t: cwd;\n\t\t\t\tconst measurement = await runImprovementMeasurement({\n\t\t\t\t\texec,\n\t\t\t\t\tcwd: measurementCwd,\n\t\t\t\t\tcommand: requireString(params.command, \"command\"),\n\t\t\t\t\tmetricName: before.config.metricName,\n\t\t\t\t\tchecksCommand: params.checksCommand,\n\t\t\t\t\ttimeoutSeconds: params.timeoutSeconds,\n\t\t\t\t\tchecksTimeoutSeconds: params.checksTimeoutSeconds,\n\t\t\t\t\tmaxOutputBytes: params.maxOutputBytes,\n\t\t\t\t\tsignal,\n\t\t\t\t});\n\t\t\t\tconst state = await recordImprovementRun({\n\t\t\t\t\tcwd,\n\t\t\t\t\tloopId: params.loopId,\n\t\t\t\t\tmetric: measurement.primaryMetric,\n\t\t\t\t\tsecondaryMetrics: secondaryMetricsFrom(measurement.parsedMetrics, before.config.metricName),\n\t\t\t\t\tchecksPass: measurement.checksPass,\n\t\t\t\t\thypothesis: params.hypothesis,\n\t\t\t\t\tchangedFiles: params.changedFiles,\n\t\t\t\t\tevidenceRef: params.evidenceRef ?? `command:${measurement.command};cwd:${measurementCwd}`,\n\t\t\t\t\tnextHint: params.nextHint,\n\t\t\t\t});\n\t\t\t\tconst decision = state.lastDecision;\n\t\t\t\treturn {\n\t\t\t\t\tcontent: [\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: `Measured ${before.config.metricName}=${measurement.primaryMetric ?? \"missing\"}; decision: ${decision?.decision ?? \"unknown\"} (${decision?.reason ?? \"unknown\"}).\\n${summarizeLoopState(state)}`,\n\t\t\t\t\t\t},\n\t\t\t\t\t],\n\t\t\t\t\tdetails: {\n\t\t\t\t\t\taction: \"measure\",\n\t\t\t\t\t\tstate,\n\t\t\t\t\t\tdecision: decision ?? undefined,\n\t\t\t\t\t\tmeasurement,\n\t\t\t\t\t\tlogPath: state.logPath,\n\t\t\t\t\t},\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tif (params.action === \"sandbox_create\") {\n\t\t\t\tif (!exec) throw new Error(\"improvement_loop sandbox_create requires createImprovementLoopTool(pi.exec)\");\n\t\t\t\tconst state = await createImprovementSandbox({\n\t\t\t\t\tcwd,\n\t\t\t\t\tloopId: params.loopId,\n\t\t\t\t\texec,\n\t\t\t\t\tbaseRef: params.baseRef,\n\t\t\t\t\tallowDirtyRepo: params.allowDirtyRepo,\n\t\t\t\t\tsandboxId: params.sandboxId,\n\t\t\t\t\tsignal,\n\t\t\t\t});\n\t\t\t\tconst sandbox = state.activeSandbox ?? undefined;\n\t\t\t\treturn {\n\t\t\t\t\tcontent: [\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: `Created sandbox: ${sandbox?.worktreePath ?? \"unknown\"}\\n${summarizeLoopState(state)}`,\n\t\t\t\t\t\t},\n\t\t\t\t\t],\n\t\t\t\t\tdetails: { action: \"sandbox_create\", state, sandbox, logPath: state.logPath },\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tif (params.action === \"sandbox_export\") {\n\t\t\t\tif (!exec) throw new Error(\"improvement_loop sandbox_export requires createImprovementLoopTool(pi.exec)\");\n\t\t\t\tconst state = await exportImprovementSandboxPatch({\n\t\t\t\t\tcwd,\n\t\t\t\t\tloopId: params.loopId,\n\t\t\t\t\texec,\n\t\t\t\t\tsandboxId: params.sandboxId,\n\t\t\t\t\tallowEmptyPatch: params.allowEmptyPatch,\n\t\t\t\t\tsignal,\n\t\t\t\t});\n\t\t\t\tconst sandbox = params.sandboxId\n\t\t\t\t\t? state.sandboxes.find((candidate) => candidate.sandboxId === normalizeLoopId(params.sandboxId))\n\t\t\t\t\t: (state.activeSandbox ?? state.sandboxes.at(-1));\n\t\t\t\treturn {\n\t\t\t\t\tcontent: [\n\t\t\t\t\t\t{\n\t\t\t\t\t\t\ttype: \"text\",\n\t\t\t\t\t\t\ttext: `Exported sandbox patch: ${sandbox?.patchPath ?? \"unknown\"}\\n${summarizeLoopState(state)}`,\n\t\t\t\t\t\t},\n\t\t\t\t\t],\n\t\t\t\t\tdetails: { action: \"sandbox_export\", state, sandbox, logPath: state.logPath },\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tif (params.action === \"sandbox_cleanup\") {\n\t\t\t\tif (!exec) throw new Error(\"improvement_loop sandbox_cleanup requires createImprovementLoopTool(pi.exec)\");\n\t\t\t\tconst state = await cleanupImprovementSandbox({\n\t\t\t\t\tcwd,\n\t\t\t\t\tloopId: params.loopId,\n\t\t\t\t\texec,\n\t\t\t\t\tsandboxId: params.sandboxId,\n\t\t\t\t\treason: params.cleanupReason,\n\t\t\t\t\tsignal,\n\t\t\t\t});\n\t\t\t\tconst sandbox = params.sandboxId\n\t\t\t\t\t? state.sandboxes.find((candidate) => candidate.sandboxId === normalizeLoopId(params.sandboxId))\n\t\t\t\t\t: state.sandboxes.at(-1);\n\t\t\t\treturn {\n\t\t\t\t\tcontent: [{ type: \"text\", text: `Cleaned sandbox.\\n${summarizeLoopState(state)}` }],\n\t\t\t\t\tdetails: { action: \"sandbox_cleanup\", state, sandbox, logPath: state.logPath },\n\t\t\t\t};\n\t\t\t}\n\n\t\t\tconst state = await readImprovementLoopState({ cwd, loopId: params.loopId });\n\t\t\tconst paths = improvementLoopPaths({ cwd, loopId: params.loopId });\n\t\t\treturn {\n\t\t\t\tcontent: [{ type: \"text\", text: summarizeLoopState(state) }],\n\t\t\t\tdetails: { action: \"status\", state, logPath: state?.logPath ?? paths.logPath },\n\t\t\t};\n\t\t},\n\t});\n}\n"]}