darwin-agents 0.4.4

package/dist/evolution/safety.d.ts

@@ -0,0 +1,76 @@
+/**
+ * Darwin — Safety Gate
+ *
+ * Guards against regressions during prompt evolution.
+ * Enforces minimum data requirements, regression checks,
+ * rollback triggers, and A/B test evaluation rules.
+ */
+import type { PromptVersionStats, SafetyThresholds, DarwinExperiment } from '../types.js';
+export type ABTestOutcome = 'a_wins' | 'b_wins' | 'continue';
+export interface ABTestConfidence {
+    /** Effect size (Cohen's d approximation) */
+    effectSize: number;
+    /** Whether the result meets minimum confidence threshold */
+    confident: boolean;
+}
+export declare class SafetyGate {
+    private thresholds;
+    constructor(thresholds?: SafetyThresholds);
+    /**
+     * Check whether an agent has accumulated enough data points
+     * to proceed with evolution (prompt optimization).
+     */
+    canEvolve(_agentName: string, stats: PromptVersionStats): boolean;
+    /**
+     * Check whether score B is NOT a regression beyond the allowed threshold.
+     *
+     * Returns `true` if B is acceptable (no regression or within tolerance).
+     * Returns `false` if B has regressed beyond `maxRegression` compared to A.
+     *
+     * Example: maxRegression = 0.20, scoreA = 0.80
+     *   - scoreB = 0.70 => drop = 0.125 (12.5%) => acceptable
+     *   - scoreB = 0.60 => drop = 0.250 (25.0%) => regression
+     */
+    checkRegression(scoreA: number, scoreB: number): boolean;
+    /**
+     * Check if the agent should roll back to its last-known-good prompt
+     * based on consecutive failure count.
+     */
+    shouldRollback(consecutiveFailures: number): boolean;
+    /**
+     * Evaluate the outcome of an A/B test between two prompt versions.
+     *
+     * Rules:
+     *   1. Both versions need at least `minRuns` total attempts (success + fail).
+     *   2. If a version has >50% failure rate with 3+ attempts, it auto-loses.
+     *   3. The winner must show >5% improvement in composite score.
+     *   4. If neither clears the bar, the test continues.
+     *
+     * @param overrideMinRuns — Per-test minimum runs (from ABTest.minRuns).
+     *   Falls back to SafetyThresholds.minDataPoints if not provided.
+     */
+    evaluateABTest(compositeA: number, compositeB: number, runsA: number, runsB: number, failsA?: number, failsB?: number, overrideMinRuns?: number): ABTestOutcome;
+    /**
+     * Calculate a simple confidence metric for an A/B test result.
+     * Uses effect size (difference / pooled estimate) as a proxy.
+     * Minimum sample: both sides need >= minDataPoints runs.
+     */
+    calculateConfidence(compositeA: number, compositeB: number, runsA: number, runsB: number): ABTestConfidence;
+    /**
+     * Compute dynamic minRuns based on observed quality score variance.
+     *
+     * When scores cluster tightly (e.g., 6.5-7.0, std < 0.5), a small
+     * sample cannot distinguish between versions. This method increases
+     * minRuns proportionally to the inverse of variance:
+     *
+     *   - High variance (std >= 1.0): floor (10) — easy to detect differences
+     *   - Low variance  (std < 0.5):  ceil (30) — need more samples
+     *   - Mid variance  (0.5 - 1.0):  linear interpolation
+     *
+     * @param experiments — Recent experiments for both A and B versions
+     * @param configMinRuns — Agent-level minRuns override from EvolutionConfig
+     * @returns Computed minRuns (never below floor, never above ceil)
+     */
+    computeDynamicMinRuns(experiments: DarwinExperiment[], configMinRuns?: number): number;
+}
+//# sourceMappingURL=safety.d.ts.map

package/dist/evolution/safety.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"safety.d.ts","sourceRoot":"","sources":["../../src/evolution/safety.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAEH,OAAO,KAAK,EAAE,kBAAkB,EAAE,gBAAgB,EAAE,gBAAgB,EAAE,MAAM,aAAa,CAAC;AAG1F,MAAM,MAAM,aAAa,GAAG,QAAQ,GAAG,QAAQ,GAAG,UAAU,CAAC;AAE7D,MAAM,WAAW,gBAAgB;IAC/B,4CAA4C;IAC5C,UAAU,EAAE,MAAM,CAAC;IACnB,4DAA4D;IAC5D,SAAS,EAAE,OAAO,CAAC;CACpB;AAMD,qBAAa,UAAU;IACrB,OAAO,CAAC,UAAU,CAAmB;gBAEzB,UAAU,GAAE,gBAAiC;IAIzD;;;OAGG;IACH,SAAS,CAAC,UAAU,EAAE,MAAM,EAAE,KAAK,EAAE,kBAAkB,GAAG,OAAO;IAIjE;;;;;;;;;OASG;IACH,eAAe,CAAC,MAAM,EAAE,MAAM,EAAE,MAAM,EAAE,MAAM,GAAG,OAAO;IAUxD;;;OAGG;IACH,cAAc,CAAC,mBAAmB,EAAE,MAAM,GAAG,OAAO;IAIpD;;;;;;;;;;;OAWG;IACH,cAAc,CACZ,UAAU,EAAE,MAAM,EAClB,UAAU,EAAE,MAAM,EAClB,KAAK,EAAE,MAAM,EACb,KAAK,EAAE,MAAM,EACb,MAAM,GAAE,MAAU,EAClB,MAAM,GAAE,MAAU,EAClB,eAAe,CAAC,EAAE,MAAM,GACvB,aAAa;IA+DhB;;;;OAIG;IACH,mBAAmB,CACjB,UAAU,EAAE,MAAM,EAClB,UAAU,EAAE,MAAM,EAClB,KAAK,EAAE,MAAM,EACb,KAAK,EAAE,MAAM,GACZ,gBAAgB;IAuBnB;;;;;;;;;;;;;;OAcG;IACH,qBAAqB,CACnB,WAAW,EAAE,gBAAgB,EAAE,EAC/B,aAAa,CAAC,EAAE,MAAM,GACrB,MAAM;CAkCV"}

package/dist/evolution/safety.js

@@ -0,0 +1,182 @@
+/**
+ * Darwin — Safety Gate
+ *
+ * Guards against regressions during prompt evolution.
+ * Enforces minimum data requirements, regression checks,
+ * rollback triggers, and A/B test evaluation rules.
+ */
+import { DEFAULT_SAFETY } from '../types.js';
+/** Default minRuns range for dynamic sizing */
+const DYNAMIC_MIN_RUNS_FLOOR = 10;
+const DYNAMIC_MIN_RUNS_CEIL = 30;
+export class SafetyGate {
+    thresholds;
+    constructor(thresholds = DEFAULT_SAFETY) {
+        this.thresholds = thresholds;
+    }
+    /**
+     * Check whether an agent has accumulated enough data points
+     * to proceed with evolution (prompt optimization).
+     */
+    canEvolve(_agentName, stats) {
+        return stats.totalRuns >= this.thresholds.minDataPoints;
+    }
+    /**
+     * Check whether score B is NOT a regression beyond the allowed threshold.
+     *
+     * Returns `true` if B is acceptable (no regression or within tolerance).
+     * Returns `false` if B has regressed beyond `maxRegression` compared to A.
+     *
+     * Example: maxRegression = 0.20, scoreA = 0.80
+     *   - scoreB = 0.70 => drop = 0.125 (12.5%) => acceptable
+     *   - scoreB = 0.60 => drop = 0.250 (25.0%) => regression
+     */
+    checkRegression(scoreA, scoreB) {
+        // If A is zero or negative, any B is acceptable (no baseline)
+        if (scoreA <= 0) {
+            return true;
+        }
+        const drop = (scoreA - scoreB) / scoreA;
+        return drop <= this.thresholds.maxRegression;
+    }
+    /**
+     * Check if the agent should roll back to its last-known-good prompt
+     * based on consecutive failure count.
+     */
+    shouldRollback(consecutiveFailures) {
+        return consecutiveFailures >= this.thresholds.failureRollbackThreshold;
+    }
+    /**
+     * Evaluate the outcome of an A/B test between two prompt versions.
+     *
+     * Rules:
+     *   1. Both versions need at least `minRuns` total attempts (success + fail).
+     *   2. If a version has >50% failure rate with 3+ attempts, it auto-loses.
+     *   3. The winner must show >5% improvement in composite score.
+     *   4. If neither clears the bar, the test continues.
+     *
+     * @param overrideMinRuns — Per-test minimum runs (from ABTest.minRuns).
+     *   Falls back to SafetyThresholds.minDataPoints if not provided.
+     */
+    evaluateABTest(compositeA, compositeB, runsA, runsB, failsA = 0, failsB = 0, overrideMinRuns) {
+        const minRuns = overrideMinRuns ?? this.thresholds.minDataPoints;
+        const totalA = runsA + failsA;
+        const totalB = runsB + failsB;
+        // Reliability check: if a version fails >50% with 3+ total attempts, it auto-loses
+        const minAttemptsForReliability = 3;
+        if (totalB >= minAttemptsForReliability && failsB / totalB > 0.5) {
+            return 'a_wins'; // B is unreliable
+        }
+        if (totalA >= minAttemptsForReliability && failsA / totalA > 0.5) {
+            return 'b_wins'; // A is unreliable
+        }
+        // Not enough successful data on either side — keep testing
+        if (runsA < minRuns || runsB < minRuns) {
+            return 'continue';
+        }
+        const improvementThreshold = 0.05; // 5% relative improvement needed
+        // Factor reliability into composite: penalize versions with failures
+        const reliabilityA = totalA > 0 ? runsA / totalA : 1;
+        const reliabilityB = totalB > 0 ? runsB / totalB : 1;
+        const adjustedA = compositeA * reliabilityA;
+        const adjustedB = compositeB * reliabilityB;
+        // Avoid division by zero
+        if (adjustedA === 0 && adjustedB === 0) {
+            return 'continue';
+        }
+        // Check if B beats A by >5%
+        if (adjustedA > 0) {
+            const bOverA = (adjustedB - adjustedA) / adjustedA;
+            if (bOverA > improvementThreshold) {
+                return 'b_wins';
+            }
+        }
+        else if (adjustedB > 0) {
+            return 'b_wins';
+        }
+        // Check if A beats B by >5%
+        if (adjustedB > 0) {
+            const aOverB = (adjustedA - adjustedB) / adjustedB;
+            if (aOverB > improvementThreshold) {
+                return 'a_wins';
+            }
+        }
+        else if (adjustedA > 0) {
+            return 'a_wins';
+        }
+        // Neither version has a decisive advantage.
+        // But prevent infinite tests: if both have 2x minRuns, declare incumbent (A) the winner.
+        // Rationale: if B can't prove itself better after double the sample, A keeps its position.
+        const maxRunsPerSide = minRuns * 2;
+        if (runsA >= maxRunsPerSide && runsB >= maxRunsPerSide) {
+            return 'a_wins'; // Incumbent wins by default — challenger failed to prove superiority
+        }
+        return 'continue';
+    }
+    /**
+     * Calculate a simple confidence metric for an A/B test result.
+     * Uses effect size (difference / pooled estimate) as a proxy.
+     * Minimum sample: both sides need >= minDataPoints runs.
+     */
+    calculateConfidence(compositeA, compositeB, runsA, runsB) {
+        const minRuns = this.thresholds.minDataPoints;
+        if (runsA < minRuns || runsB < minRuns) {
+            return { effectSize: 0, confident: false };
+        }
+        // Pooled estimate (simple average as variance proxy)
+        const pooled = (compositeA + compositeB) / 2;
+        if (pooled === 0) {
+            return { effectSize: 0, confident: false };
+        }
+        // Effect size: absolute difference normalized by pooled mean
+        const effectSize = Math.abs(compositeA - compositeB) / pooled;
+        // Require at least "small" effect size (0.2) and enough samples
+        const totalSamples = runsA + runsB;
+        const confident = effectSize >= 0.2 && totalSamples >= minRuns * 2;
+        return { effectSize, confident };
+    }
+    /**
+     * Compute dynamic minRuns based on observed quality score variance.
+     *
+     * When scores cluster tightly (e.g., 6.5-7.0, std < 0.5), a small
+     * sample cannot distinguish between versions. This method increases
+     * minRuns proportionally to the inverse of variance:
+     *
+     *   - High variance (std >= 1.0): floor (10) — easy to detect differences
+     *   - Low variance  (std < 0.5):  ceil (30) — need more samples
+     *   - Mid variance  (0.5 - 1.0):  linear interpolation
+     *
+     * @param experiments — Recent experiments for both A and B versions
+     * @param configMinRuns — Agent-level minRuns override from EvolutionConfig
+     * @returns Computed minRuns (never below floor, never above ceil)
+     */
+    computeDynamicMinRuns(experiments, configMinRuns) {
+        const floor = configMinRuns ?? DYNAMIC_MIN_RUNS_FLOOR;
+        const ceil = Math.max(floor, DYNAMIC_MIN_RUNS_CEIL);
+        // Need at least 4 quality scores to estimate variance
+        const qualityScores = experiments
+            .map((e) => e.metrics.qualityScore)
+            .filter((s) => s !== null);
+        if (qualityScores.length < 4) {
+            return floor;
+        }
+        const mean = qualityScores.reduce((a, b) => a + b, 0) / qualityScores.length;
+        // Bessel's correction (n-1): we are estimating population variance from a sample.
+        // Without this, small samples (n=4-5) underestimate std by ~13%, inflating minRuns.
+        const variance = qualityScores.reduce((sum, s) => sum + (s - mean) ** 2, 0) / (qualityScores.length - 1);
+        const std = Math.sqrt(variance);
+        // High variance (std >= 1.0): floor — differences are easy to spot
+        if (std >= 1.0) {
+            return floor;
+        }
+        // Low variance (std < 0.5): ceil — need many samples
+        if (std < 0.5) {
+            return ceil;
+        }
+        // Mid range: linear interpolation between ceil and floor
+        // std=0.5 → ceil, std=1.0 → floor
+        const t = (std - 0.5) / 0.5; // 0 at std=0.5, 1 at std=1.0
+        return Math.round(ceil + t * (floor - ceil));
+    }
+}
+//# sourceMappingURL=safety.js.map

package/dist/evolution/safety.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"safety.js","sourceRoot":"","sources":["../../src/evolution/safety.ts"],"names":[],"mappings":"AAAA;;;;;;GAMG;AAGH,OAAO,EAAE,cAAc,EAAE,MAAM,aAAa,CAAC;AAW7C,+CAA+C;AAC/C,MAAM,sBAAsB,GAAG,EAAE,CAAC;AAClC,MAAM,qBAAqB,GAAG,EAAE,CAAC;AAEjC,MAAM,OAAO,UAAU;IACb,UAAU,CAAmB;IAErC,YAAY,aAA+B,cAAc;QACvD,IAAI,CAAC,UAAU,GAAG,UAAU,CAAC;IAC/B,CAAC;IAED;;;OAGG;IACH,SAAS,CAAC,UAAkB,EAAE,KAAyB;QACrD,OAAO,KAAK,CAAC,SAAS,IAAI,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC;IAC1D,CAAC;IAED;;;;;;;;;OASG;IACH,eAAe,CAAC,MAAc,EAAE,MAAc;QAC5C,8DAA8D;QAC9D,IAAI,MAAM,IAAI,CAAC,EAAE,CAAC;YAChB,OAAO,IAAI,CAAC;QACd,CAAC;QAED,MAAM,IAAI,GAAG,CAAC,MAAM,GAAG,MAAM,CAAC,GAAG,MAAM,CAAC;QACxC,OAAO,IAAI,IAAI,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC;IAC/C,CAAC;IAED;;;OAGG;IACH,cAAc,CAAC,mBAA2B;QACxC,OAAO,mBAAmB,IAAI,IAAI,CAAC,UAAU,CAAC,wBAAwB,CAAC;IACzE,CAAC;IAED;;;;;;;;;;;OAWG;IACH,cAAc,CACZ,UAAkB,EAClB,UAAkB,EAClB,KAAa,EACb,KAAa,EACb,SAAiB,CAAC,EAClB,SAAiB,CAAC,EAClB,eAAwB;QAExB,MAAM,OAAO,GAAG,eAAe,IAAI,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC;QACjE,MAAM,MAAM,GAAG,KAAK,GAAG,MAAM,CAAC;QAC9B,MAAM,MAAM,GAAG,KAAK,GAAG,MAAM,CAAC;QAE9B,mFAAmF;QACnF,MAAM,yBAAyB,GAAG,CAAC,CAAC;QACpC,IAAI,MAAM,IAAI,yBAAyB,IAAI,MAAM,GAAG,MAAM,GAAG,GAAG,EAAE,CAAC;YACjE,OAAO,QAAQ,CAAC,CAAC,kBAAkB;QACrC,CAAC;QACD,IAAI,MAAM,IAAI,yBAAyB,IAAI,MAAM,GAAG,MAAM,GAAG,GAAG,EAAE,CAAC;YACjE,OAAO,QAAQ,CAAC,CAAC,kBAAkB;QACrC,CAAC;QAED,2DAA2D;QAC3D,IAAI,KAAK,GAAG,OAAO,IAAI,KAAK,GAAG,OAAO,EAAE,CAAC;YACvC,OAAO,UAAU,CAAC;QACpB,CAAC;QAED,MAAM,oBAAoB,GAAG,IAAI,CAAC,CAAC,iCAAiC;QAEpE,qEAAqE;QACrE,MAAM,YAAY,GAAG,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;QACrD,MAAM,YAAY,GAAG,MAAM,GAAG,CAAC,CAAC,CAAC,CAAC,KAAK,GAAG,MAAM,CAAC,CAAC,CAAC,CAAC,CAAC;QACrD,MAAM,SAAS,GAAG,UAAU,GAAG,YAAY,CAAC;QAC5C,MAAM,SAAS,GAAG,UAAU,GAAG,YAAY,CAAC;QAE5C,yBAAyB;QACzB,IAAI,SAAS,KAAK,CAAC,IAAI,SAAS,KAAK,CAAC,EAAE,CAAC;YACvC,OAAO,UAAU,CAAC;QACpB,CAAC;QAED,4BAA4B;QAC5B,IAAI,SAAS,GAAG,CAAC,EAAE,CAAC;YAClB,MAAM,MAAM,GAAG,CAAC,SAAS,GAAG,SAAS,CAAC,GAAG,SAAS,CAAC;YACnD,IAAI,MAAM,GAAG,oBAAoB,EAAE,CAAC;gBAClC,OAAO,QAAQ,CAAC;YAClB,CAAC;QACH,CAAC;aAAM,IAAI,SAAS,GAAG,CAAC,EAAE,CAAC;YACzB,OAAO,QAAQ,CAAC;QAClB,CAAC;QAED,4BAA4B;QAC5B,IAAI,SAAS,GAAG,CAAC,EAAE,CAAC;YAClB,MAAM,MAAM,GAAG,CAAC,SAAS,GAAG,SAAS,CAAC,GAAG,SAAS,CAAC;YACnD,IAAI,MAAM,GAAG,oBAAoB,EAAE,CAAC;gBAClC,OAAO,QAAQ,CAAC;YAClB,CAAC;QACH,CAAC;aAAM,IAAI,SAAS,GAAG,CAAC,EAAE,CAAC;YACzB,OAAO,QAAQ,CAAC;QAClB,CAAC;QAED,4CAA4C;QAC5C,yFAAyF;QACzF,2FAA2F;QAC3F,MAAM,cAAc,GAAG,OAAO,GAAG,CAAC,CAAC;QACnC,IAAI,KAAK,IAAI,cAAc,IAAI,KAAK,IAAI,cAAc,EAAE,CAAC;YACvD,OAAO,QAAQ,CAAC,CAAC,qEAAqE;QACxF,CAAC;QAED,OAAO,UAAU,CAAC;IACpB,CAAC;IAED;;;;OAIG;IACH,mBAAmB,CACjB,UAAkB,EAClB,UAAkB,EAClB,KAAa,EACb,KAAa;QAEb,MAAM,OAAO,GAAG,IAAI,CAAC,UAAU,CAAC,aAAa,CAAC;QAE9C,IAAI,KAAK,GAAG,OAAO,IAAI,KAAK,GAAG,OAAO,EAAE,CAAC;YACvC,OAAO,EAAE,UAAU,EAAE,CAAC,EAAE,SAAS,EAAE,KAAK,EAAE,CAAC;QAC7C,CAAC;QAED,qDAAqD;QACrD,MAAM,MAAM,GAAG,CAAC,UAAU,GAAG,UAAU,CAAC,GAAG,CAAC,CAAC;QAC7C,IAAI,MAAM,KAAK,CAAC,EAAE,CAAC;YACjB,OAAO,EAAE,UAAU,EAAE,CAAC,EAAE,SAAS,EAAE,KAAK,EAAE,CAAC;QAC7C,CAAC;QAED,6DAA6D;QAC7D,MAAM,UAAU,GAAG,IAAI,CAAC,GAAG,CAAC,UAAU,GAAG,UAAU,CAAC,GAAG,MAAM,CAAC;QAE9D,gEAAgE;QAChE,MAAM,YAAY,GAAG,KAAK,GAAG,KAAK,CAAC;QACnC,MAAM,SAAS,GAAG,UAAU,IAAI,GAAG,IAAI,YAAY,IAAI,OAAO,GAAG,CAAC,CAAC;QAEnE,OAAO,EAAE,UAAU,EAAE,SAAS,EAAE,CAAC;IACnC,CAAC;IAED;;;;;;;;;;;;;;OAcG;IACH,qBAAqB,CACnB,WAA+B,EAC/B,aAAsB;QAEtB,MAAM,KAAK,GAAG,aAAa,IAAI,sBAAsB,CAAC;QACtD,MAAM,IAAI,GAAG,IAAI,CAAC,GAAG,CAAC,KAAK,EAAE,qBAAqB,CAAC,CAAC;QAEpD,sDAAsD;QACtD,MAAM,aAAa,GAAG,WAAW;aAC9B,GAAG,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,CAAC,YAAY,CAAC;aAClC,MAAM,CAAC,CAAC,CAAC,EAAe,EAAE,CAAC,CAAC,KAAK,IAAI,CAAC,CAAC;QAE1C,IAAI,aAAa,CAAC,MAAM,GAAG,CAAC,EAAE,CAAC;YAC7B,OAAO,KAAK,CAAC;QACf,CAAC;QAED,MAAM,IAAI,GAAG,aAAa,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,EAAE,CAAC,CAAC,GAAG,aAAa,CAAC,MAAM,CAAC;QAC7E,kFAAkF;QAClF,oFAAoF;QACpF,MAAM,QAAQ,GAAG,aAAa,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,CAAC,CAAC,GAAG,IAAI,CAAC,IAAI,CAAC,EAAE,CAAC,CAAC,GAAG,CAAC,aAAa,CAAC,MAAM,GAAG,CAAC,CAAC,CAAC;QACzG,MAAM,GAAG,GAAG,IAAI,CAAC,IAAI,CAAC,QAAQ,CAAC,CAAC;QAEhC,mEAAmE;QACnE,IAAI,GAAG,IAAI,GAAG,EAAE,CAAC;YACf,OAAO,KAAK,CAAC;QACf,CAAC;QAED,qDAAqD;QACrD,IAAI,GAAG,GAAG,GAAG,EAAE,CAAC;YACd,OAAO,IAAI,CAAC;QACd,CAAC;QAED,yDAAyD;QACzD,kCAAkC;QAClC,MAAM,CAAC,GAAG,CAAC,GAAG,GAAG,GAAG,CAAC,GAAG,GAAG,CAAC,CAAC,6BAA6B;QAC1D,OAAO,IAAI,CAAC,KAAK,CAAC,IAAI,GAAG,CAAC,GAAG,CAAC,KAAK,GAAG,IAAI,CAAC,CAAC,CAAC;IAC/C,CAAC;CACF"}

package/dist/evolution/tracker.d.ts

@@ -0,0 +1,48 @@
+/**
+ * Darwin — Experiment Tracker
+ *
+ * Records experiments, aggregates stats, and computes composite scores
+ * for prompt version evaluation.
+ */
+import type { DarwinExperiment, MemoryProvider, MetricWeights, PromptVersionStats } from '../types.js';
+import type { CategoryStats } from './optimizer.js';
+export declare class ExperimentTracker {
+    private memory;
+    constructor(memory: MemoryProvider);
+    /**
+     * Record a completed experiment.
+     * Saves it to memory, updates the prompt version stats, and adjusts
+     * the consecutive-failure counter in Darwin state.
+     */
+    recordExperiment(exp: DarwinExperiment): Promise<void>;
+    /**
+     * Aggregate stats from all experiments for a given agent (optionally
+     * filtered to a single prompt version).
+     */
+    getStats(agentName: string, version?: string): Promise<PromptVersionStats>;
+    /**
+     * Calculate a composite score for a single experiment.
+     *
+     * Normalization ranges:
+     *   quality      — score / 10          (0-10 scale)
+     *   sourceCount  — min(count / 20, 1)  (20 sources = perfect)
+     *   outputLength — min(len / 10000, 1) (10k chars = perfect)
+     *   duration     — 1 - min(ms/300000, 1) (lower is better, 5 min cap)
+     *   success      — 1 if true, 0 if false
+     */
+    getCompositeScore(exp: DarwinExperiment, weights?: MetricWeights): number;
+    /**
+     * Get stats broken down by task category (P2-5).
+     * Gives the optimizer visibility into which topic types perform well/poorly.
+     */
+    getStatsByCategory(agentName: string): Promise<CategoryStats[]>;
+    /**
+     * Average composite score across experiments for a specific agent + prompt version.
+     *
+     * If `since` is provided, only experiments after that ISO timestamp are included.
+     * This is critical for A/B tests: compare only the test period, not all-time data
+     * (otherwise the incumbent version's historical data skews the comparison).
+     */
+    getAverageComposite(agentName: string, version: string, weights?: MetricWeights, since?: string): Promise<number>;
+}
+//# sourceMappingURL=tracker.d.ts.map

package/dist/evolution/tracker.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"tracker.d.ts","sourceRoot":"","sources":["../../src/evolution/tracker.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EACV,gBAAgB,EAEhB,cAAc,EACd,aAAa,EACb,kBAAkB,EACnB,MAAM,aAAa,CAAC;AACrB,OAAO,KAAK,EAAE,aAAa,EAAE,MAAM,gBAAgB,CAAC;AAGpD,qBAAa,iBAAiB;IAC5B,OAAO,CAAC,MAAM,CAAiB;gBAEnB,MAAM,EAAE,cAAc;IAIlC;;;;OAIG;IACG,gBAAgB,CAAC,GAAG,EAAE,gBAAgB,GAAG,OAAO,CAAC,IAAI,CAAC;IA+B5D;;;OAGG;IACG,QAAQ,CACZ,SAAS,EAAE,MAAM,EACjB,OAAO,CAAC,EAAE,MAAM,GACf,OAAO,CAAC,kBAAkB,CAAC;IA6C9B;;;;;;;;;OASG;IACH,iBAAiB,CACf,GAAG,EAAE,gBAAgB,EACrB,OAAO,GAAE,aAA+B,GACvC,MAAM;IA8BT;;;OAGG;IACG,kBAAkB,CAAC,SAAS,EAAE,MAAM,GAAG,OAAO,CAAC,aAAa,EAAE,CAAC;IA6BrE;;;;;;OAMG;IACG,mBAAmB,CACvB,SAAS,EAAE,MAAM,EACjB,OAAO,EAAE,MAAM,EACf,OAAO,GAAE,aAA+B,EACxC,KAAK,CAAC,EAAE,MAAM,GACb,OAAO,CAAC,MAAM,CAAC;CAmBnB"}

package/dist/evolution/tracker.js

@@ -0,0 +1,163 @@
+/**
+ * Darwin — Experiment Tracker
+ *
+ * Records experiments, aggregates stats, and computes composite scores
+ * for prompt version evaluation.
+ */
+import { DEFAULT_WEIGHTS } from '../types.js';
+export class ExperimentTracker {
+    memory;
+    constructor(memory) {
+        this.memory = memory;
+    }
+    /**
+     * Record a completed experiment.
+     * Saves it to memory, updates the prompt version stats, and adjusts
+     * the consecutive-failure counter in Darwin state.
+     */
+    async recordExperiment(exp) {
+        // 1. Persist the raw experiment
+        await this.memory.saveExperiment(exp);
+        // 2. Refresh aggregated stats on the prompt version
+        const versions = await this.memory.getAllPromptVersions(exp.agentName);
+        const version = versions.find((v) => v.version === exp.promptVersion);
+        if (version) {
+            const updatedStats = await this.getStats(exp.agentName, exp.promptVersion);
+            version.stats = updatedStats;
+            await this.memory.savePromptVersion(version);
+        }
+        // 3. Atomically update Darwin state (experiment count + consecutive failures).
+        // Uses updateState() to prevent race conditions when multiple agents
+        // record experiments concurrently — getState()+saveState() would lose updates.
+        await this.memory.updateState((state) => {
+            state.experimentCounts[exp.agentName] =
+                (state.experimentCounts[exp.agentName] ?? 0) + 1;
+            if (exp.success) {
+                state.consecutiveFailures[exp.agentName] = 0;
+            }
+            else {
+                state.consecutiveFailures[exp.agentName] =
+                    (state.consecutiveFailures[exp.agentName] ?? 0) + 1;
+            }
+            return state;
+        });
+    }
+    /**
+     * Aggregate stats from all experiments for a given agent (optionally
+     * filtered to a single prompt version).
+     */
+    async getStats(agentName, version) {
+        const experiments = await this.memory.loadExperiments(agentName);
+        const filtered = version
+            ? experiments.filter((e) => e.promptVersion === version)
+            : experiments;
+        if (filtered.length === 0) {
+            return {
+                totalRuns: 0,
+                avgQuality: 0,
+                avgDuration: 0,
+                successRate: 0,
+                avgSourceCount: 0,
+            };
+        }
+        const totalRuns = filtered.length;
+        const successCount = filtered.filter((e) => e.success).length;
+        // Quality: only count experiments that have a quality score
+        const withQuality = filtered.filter((e) => e.metrics.qualityScore !== null);
+        const avgQuality = withQuality.length > 0
+            ? withQuality.reduce((sum, e) => sum + (e.metrics.qualityScore ?? 0), 0) /
+                withQuality.length
+            : 0;
+        const avgDuration = filtered.reduce((sum, e) => sum + e.metrics.durationMs, 0) / totalRuns;
+        const avgSourceCount = filtered.reduce((sum, e) => sum + e.metrics.sourceCount, 0) / totalRuns;
+        return {
+            totalRuns,
+            avgQuality,
+            avgDuration,
+            successRate: successCount / totalRuns,
+            avgSourceCount,
+        };
+    }
+    /**
+     * Calculate a composite score for a single experiment.
+     *
+     * Normalization ranges:
+     *   quality      — score / 10          (0-10 scale)
+     *   sourceCount  — min(count / 20, 1)  (20 sources = perfect)
+     *   outputLength — min(len / 10000, 1) (10k chars = perfect)
+     *   duration     — 1 - min(ms/300000, 1) (lower is better, 5 min cap)
+     *   success      — 1 if true, 0 if false
+     */
+    getCompositeScore(exp, weights = DEFAULT_WEIGHTS) {
+        // NULL quality = critic failed, not agent failed. Exclude from quality component
+        // instead of treating as 0 (which unfairly tanks the composite).
+        const hasQuality = exp.metrics.qualityScore !== null;
+        const qualityNorm = hasQuality ? exp.metrics.qualityScore / 10 : 0;
+        // If no quality score, redistribute quality weight to other metrics
+        const effectiveQualityWeight = hasQuality ? weights.quality : 0;
+        const weightSum = effectiveQualityWeight + weights.sourceCount + weights.outputLength + weights.duration + weights.success;
+        const scale = weightSum > 0 ? 1 / weightSum : 0;
+        const normalized = {
+            quality: qualityNorm,
+            sourceCount: Math.min(exp.metrics.sourceCount / 20, 1),
+            outputLength: Math.min(exp.metrics.outputLength / 10000, 1),
+            duration: 1 - Math.min(exp.metrics.durationMs / 300000, 1),
+            success: exp.success ? 1 : 0,
+        };
+        const score = (normalized.quality * effectiveQualityWeight +
+            normalized.sourceCount * weights.sourceCount +
+            normalized.outputLength * weights.outputLength +
+            normalized.duration * weights.duration +
+            normalized.success * weights.success) * scale;
+        return score;
+    }
+    /**
+     * Get stats broken down by task category (P2-5).
+     * Gives the optimizer visibility into which topic types perform well/poorly.
+     */
+    async getStatsByCategory(agentName) {
+        const experiments = await this.memory.loadExperiments(agentName);
+        const byCategory = new Map();
+        for (const exp of experiments) {
+            const key = exp.taskType || 'general';
+            const list = byCategory.get(key);
+            if (list) {
+                list.push(exp);
+            }
+            else {
+                byCategory.set(key, [exp]);
+            }
+        }
+        const result = [];
+        for (const [taskType, exps] of byCategory) {
+            const withQuality = exps.filter((e) => e.metrics.qualityScore !== null);
+            const avgQuality = withQuality.length > 0
+                ? withQuality.reduce((s, e) => s + (e.metrics.qualityScore ?? 0), 0) / withQuality.length
+                : 0;
+            const avgSourceCount = exps.reduce((s, e) => s + e.metrics.sourceCount, 0) / exps.length;
+            const successRate = exps.filter((e) => e.success).length / exps.length;
+            result.push({ taskType, totalRuns: exps.length, avgQuality, avgSourceCount, successRate });
+        }
+        return result.sort((a, b) => b.totalRuns - a.totalRuns);
+    }
+    /**
+     * Average composite score across experiments for a specific agent + prompt version.
+     *
+     * If `since` is provided, only experiments after that ISO timestamp are included.
+     * This is critical for A/B tests: compare only the test period, not all-time data
+     * (otherwise the incumbent version's historical data skews the comparison).
+     */
+    async getAverageComposite(agentName, version, weights = DEFAULT_WEIGHTS, since) {
+        const experiments = await this.memory.loadExperiments(agentName);
+        let filtered = experiments.filter((e) => e.promptVersion === version);
+        if (since) {
+            filtered = filtered.filter((e) => e.startedAt >= since);
+        }
+        if (filtered.length === 0) {
+            return 0;
+        }
+        const total = filtered.reduce((sum, exp) => sum + this.getCompositeScore(exp, weights), 0);
+        return total / filtered.length;
+    }
+}
+//# sourceMappingURL=tracker.js.map

package/dist/evolution/tracker.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"tracker.js","sourceRoot":"","sources":["../../src/evolution/tracker.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAUH,OAAO,EAAE,eAAe,EAAE,MAAM,aAAa,CAAC;AAE9C,MAAM,OAAO,iBAAiB;IACpB,MAAM,CAAiB;IAE/B,YAAY,MAAsB;QAChC,IAAI,CAAC,MAAM,GAAG,MAAM,CAAC;IACvB,CAAC;IAED;;;;OAIG;IACH,KAAK,CAAC,gBAAgB,CAAC,GAAqB;QAC1C,gCAAgC;QAChC,MAAM,IAAI,CAAC,MAAM,CAAC,cAAc,CAAC,GAAG,CAAC,CAAC;QAEtC,oDAAoD;QACpD,MAAM,QAAQ,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,oBAAoB,CAAC,GAAG,CAAC,SAAS,CAAC,CAAC;QACvE,MAAM,OAAO,GAAG,QAAQ,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,KAAK,GAAG,CAAC,aAAa,CAAC,CAAC;QACtE,IAAI,OAAO,EAAE,CAAC;YACZ,MAAM,YAAY,GAAG,MAAM,IAAI,CAAC,QAAQ,CAAC,GAAG,CAAC,SAAS,EAAE,GAAG,CAAC,aAAa,CAAC,CAAC;YAC3E,OAAO,CAAC,KAAK,GAAG,YAAY,CAAC;YAC7B,MAAM,IAAI,CAAC,MAAM,CAAC,iBAAiB,CAAC,OAAO,CAAC,CAAC;QAC/C,CAAC;QAED,+EAA+E;QAC/E,qEAAqE;QACrE,+EAA+E;QAC/E,MAAM,IAAI,CAAC,MAAM,CAAC,WAAW,CAAC,CAAC,KAAK,EAAE,EAAE;YACtC,KAAK,CAAC,gBAAgB,CAAC,GAAG,CAAC,SAAS,CAAC;gBACnC,CAAC,KAAK,CAAC,gBAAgB,CAAC,GAAG,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC;YAEnD,IAAI,GAAG,CAAC,OAAO,EAAE,CAAC;gBAChB,KAAK,CAAC,mBAAmB,CAAC,GAAG,CAAC,SAAS,CAAC,GAAG,CAAC,CAAC;YAC/C,CAAC;iBAAM,CAAC;gBACN,KAAK,CAAC,mBAAmB,CAAC,GAAG,CAAC,SAAS,CAAC;oBACtC,CAAC,KAAK,CAAC,mBAAmB,CAAC,GAAG,CAAC,SAAS,CAAC,IAAI,CAAC,CAAC,GAAG,CAAC,CAAC;YACxD,CAAC;YAED,OAAO,KAAK,CAAC;QACf,CAAC,CAAC,CAAC;IACL,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,QAAQ,CACZ,SAAiB,EACjB,OAAgB;QAEhB,MAAM,WAAW,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,eAAe,CAAC,SAAS,CAAC,CAAC;QAEjE,MAAM,QAAQ,GAAG,OAAO;YACtB,CAAC,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,aAAa,KAAK,OAAO,CAAC;YACxD,CAAC,CAAC,WAAW,CAAC;QAEhB,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAC1B,OAAO;gBACL,SAAS,EAAE,CAAC;gBACZ,UAAU,EAAE,CAAC;gBACb,WAAW,EAAE,CAAC;gBACd,WAAW,EAAE,CAAC;gBACd,cAAc,EAAE,CAAC;aAClB,CAAC;QACJ,CAAC;QAED,MAAM,SAAS,GAAG,QAAQ,CAAC,MAAM,CAAC;QAClC,MAAM,YAAY,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,MAAM,CAAC;QAE9D,4DAA4D;QAC5D,MAAM,WAAW,GAAG,QAAQ,CAAC,MAAM,CACjC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,CAAC,YAAY,KAAK,IAAI,CACvC,CAAC;QACF,MAAM,UAAU,GACd,WAAW,CAAC,MAAM,GAAG,CAAC;YACpB,CAAC,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,CAAC,CAAC,CAAC,OAAO,CAAC,YAAY,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC;gBACtE,WAAW,CAAC,MAAM;YACpB,CAAC,CAAC,CAAC,CAAC;QAER,MAAM,WAAW,GACf,QAAQ,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,CAAC,CAAC,OAAO,CAAC,UAAU,EAAE,CAAC,CAAC,GAAG,SAAS,CAAC;QAEzE,MAAM,cAAc,GAClB,QAAQ,CAAC,MAAM,CAAC,CAAC,GAAG,EAAE,CAAC,EAAE,EAAE,CAAC,GAAG,GAAG,CAAC,CAAC,OAAO,CAAC,WAAW,EAAE,CAAC,CAAC,GAAG,SAAS,CAAC;QAE1E,OAAO;YACL,SAAS;YACT,UAAU;YACV,WAAW;YACX,WAAW,EAAE,YAAY,GAAG,SAAS;YACrC,cAAc;SACf,CAAC;IACJ,CAAC;IAED;;;;;;;;;OASG;IACH,iBAAiB,CACf,GAAqB,EACrB,UAAyB,eAAe;QAExC,iFAAiF;QACjF,iEAAiE;QACjE,MAAM,UAAU,GAAG,GAAG,CAAC,OAAO,CAAC,YAAY,KAAK,IAAI,CAAC;QACrD,MAAM,WAAW,GAAG,UAAU,CAAC,CAAC,CAAE,GAAG,CAAC,OAAO,CAAC,YAAuB,GAAG,EAAE,CAAC,CAAC,CAAC,CAAC,CAAC;QAE/E,oEAAoE;QACpE,MAAM,sBAAsB,GAAG,UAAU,CAAC,CAAC,CAAC,OAAO,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC;QAChE,MAAM,SAAS,GAAG,sBAAsB,GAAG,OAAO,CAAC,WAAW,GAAG,OAAO,CAAC,YAAY,GAAG,OAAO,CAAC,QAAQ,GAAG,OAAO,CAAC,OAAO,CAAC;QAC3H,MAAM,KAAK,GAAG,SAAS,GAAG,CAAC,CAAC,CAAC,CAAC,CAAC,GAAG,SAAS,CAAC,CAAC,CAAC,CAAC,CAAC;QAEhD,MAAM,UAAU,GAAG;YACjB,OAAO,EAAE,WAAW;YACpB,WAAW,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,WAAW,GAAG,EAAE,EAAE,CAAC,CAAC;YACtD,YAAY,EAAE,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,YAAY,GAAG,KAAK,EAAE,CAAC,CAAC;YAC3D,QAAQ,EAAE,CAAC,GAAG,IAAI,CAAC,GAAG,CAAC,GAAG,CAAC,OAAO,CAAC,UAAU,GAAG,MAAM,EAAE,CAAC,CAAC;YAC1D,OAAO,EAAE,GAAG,CAAC,OAAO,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC,CAAC;SAC7B,CAAC;QAEF,MAAM,KAAK,GAAG,CACZ,UAAU,CAAC,OAAO,GAAG,sBAAsB;YAC3C,UAAU,CAAC,WAAW,GAAG,OAAO,CAAC,WAAW;YAC5C,UAAU,CAAC,YAAY,GAAG,OAAO,CAAC,YAAY;YAC9C,UAAU,CAAC,QAAQ,GAAG,OAAO,CAAC,QAAQ;YACtC,UAAU,CAAC,OAAO,GAAG,OAAO,CAAC,OAAO,CACrC,GAAG,KAAK,CAAC;QAEV,OAAO,KAAK,CAAC;IACf,CAAC;IAED;;;OAGG;IACH,KAAK,CAAC,kBAAkB,CAAC,SAAiB;QACxC,MAAM,WAAW,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,eAAe,CAAC,SAAS,CAAC,CAAC;QACjE,MAAM,UAAU,GAAG,IAAI,GAAG,EAA8B,CAAC;QAEzD,KAAK,MAAM,GAAG,IAAI,WAAW,EAAE,CAAC;YAC9B,MAAM,GAAG,GAAG,GAAG,CAAC,QAAQ,IAAI,SAAS,CAAC;YACtC,MAAM,IAAI,GAAG,UAAU,CAAC,GAAG,CAAC,GAAG,CAAC,CAAC;YACjC,IAAI,IAAI,EAAE,CAAC;gBACT,IAAI,CAAC,IAAI,CAAC,GAAG,CAAC,CAAC;YACjB,CAAC;iBAAM,CAAC;gBACN,UAAU,CAAC,GAAG,CAAC,GAAG,EAAE,CAAC,GAAG,CAAC,CAAC,CAAC;YAC7B,CAAC;QACH,CAAC;QAED,MAAM,MAAM,GAAoB,EAAE,CAAC;QACnC,KAAK,MAAM,CAAC,QAAQ,EAAE,IAAI,CAAC,IAAI,UAAU,EAAE,CAAC;YAC1C,MAAM,WAAW,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,CAAC,YAAY,KAAK,IAAI,CAAC,CAAC;YACxE,MAAM,UAAU,GAAG,WAAW,CAAC,MAAM,GAAG,CAAC;gBACvC,CAAC,CAAC,WAAW,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,CAAC,OAAO,CAAC,YAAY,IAAI,CAAC,CAAC,EAAE,CAAC,CAAC,GAAG,WAAW,CAAC,MAAM;gBACzF,CAAC,CAAC,CAAC,CAAC;YACN,MAAM,cAAc,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,GAAG,CAAC,CAAC,OAAO,CAAC,WAAW,EAAE,CAAC,CAAC,GAAG,IAAI,CAAC,MAAM,CAAC;YACzF,MAAM,WAAW,GAAG,IAAI,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,OAAO,CAAC,CAAC,MAAM,GAAG,IAAI,CAAC,MAAM,CAAC;YAEvE,MAAM,CAAC,IAAI,CAAC,EAAE,QAAQ,EAAE,SAAS,EAAE,IAAI,CAAC,MAAM,EAAE,UAAU,EAAE,cAAc,EAAE,WAAW,EAAE,CAAC,CAAC;QAC7F,CAAC;QAED,OAAO,MAAM,CAAC,IAAI,CAAC,CAAC,CAAC,EAAE,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,GAAG,CAAC,CAAC,SAAS,CAAC,CAAC;IAC1D,CAAC;IAED;;;;;;OAMG;IACH,KAAK,CAAC,mBAAmB,CACvB,SAAiB,EACjB,OAAe,EACf,UAAyB,eAAe,EACxC,KAAc;QAEd,MAAM,WAAW,GAAG,MAAM,IAAI,CAAC,MAAM,CAAC,eAAe,CAAC,SAAS,CAAC,CAAC;QACjE,IAAI,QAAQ,GAAG,WAAW,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,aAAa,KAAK,OAAO,CAAC,CAAC;QAEtE,IAAI,KAAK,EAAE,CAAC;YACV,QAAQ,GAAG,QAAQ,CAAC,MAAM,CAAC,CAAC,CAAC,EAAE,EAAE,CAAC,CAAC,CAAC,SAAS,IAAI,KAAK,CAAC,CAAC;QAC1D,CAAC;QAED,IAAI,QAAQ,CAAC,MAAM,KAAK,CAAC,EAAE,CAAC;YAC1B,OAAO,CAAC,CAAC;QACX,CAAC;QAED,MAAM,KAAK,GAAG,QAAQ,CAAC,MAAM,CAC3B,CAAC,GAAG,EAAE,GAAG,EAAE,EAAE,CAAC,GAAG,GAAG,IAAI,CAAC,iBAAiB,CAAC,GAAG,EAAE,OAAO,CAAC,EACxD,CAAC,CACF,CAAC;QAEF,OAAO,KAAK,GAAG,QAAQ,CAAC,MAAM,CAAC;IACjC,CAAC;CACF"}

package/dist/index.d.ts

@@ -0,0 +1,32 @@
+/**
+ * Darwin — AI agents that improve themselves.
+ *
+ * @example
+ * ```typescript
+ * import { defineAgent, defineConfig, runAgent } from 'darwin-agents';
+ *
+ * const myAgent = defineAgent({
+ *   name: 'summarizer',
+ *   role: 'Text Summarizer',
+ *   systemPrompt: 'Summarize text in 3 bullet points.',
+ *   evolution: { enabled: true, evaluator: 'critic' },
+ * });
+ *
+ * const result = await runAgent(myAgent, 'Summarize this article...');
+ * ```
+ */
+export { defineAgent, defineConfig, loadConfig, loadConfigSync } from './core/agent.js';
+export { runAgent } from './core/runner.js';
+export type { AgentDefinition, DarwinConfig, DarwinExperiment, DarwinMetrics, DarwinPattern, DarwinState, ExperimentFeedback, EvolutionConfig, Learning, McpServerConfig, MemoryProvider, MetricWeights, PromptVersion, PromptVersionStats, RunResult, SafetyThresholds, } from './types.js';
+export { DEFAULT_WEIGHTS, DEFAULT_SAFETY } from './types.js';
+export { writer, researcher, critic, analyst, builtinAgents } from './agents/index.js';
+export { createProvider } from './providers/index.js';
+export type { LLMProvider, LLMCallOptions, LLMCallResult, ProviderConfig } from './providers/types.js';
+export { AnthropicProvider } from './providers/anthropic.js';
+export { OpenAIProvider } from './providers/openai.js';
+export { OllamaProvider } from './providers/ollama.js';
+export { ClaudeCliProvider } from './providers/claude-cli.js';
+export { createMemory, SqliteMemoryProvider, PostgresMemoryProvider } from './memory/index.js';
+export { loadNotificationConfig } from './evolution/notifications.js';
+export type { NotificationConfig } from './evolution/notifications.js';
+//# sourceMappingURL=index.d.ts.map

package/dist/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAGH,OAAO,EAAE,WAAW,EAAE,YAAY,EAAE,UAAU,EAAE,cAAc,EAAE,MAAM,iBAAiB,CAAC;AACxF,OAAO,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AAG5C,YAAY,EACV,eAAe,EACf,YAAY,EACZ,gBAAgB,EAChB,aAAa,EACb,aAAa,EACb,WAAW,EACX,kBAAkB,EAClB,eAAe,EACf,QAAQ,EACR,eAAe,EACf,cAAc,EACd,aAAa,EACb,aAAa,EACb,kBAAkB,EAClB,SAAS,EACT,gBAAgB,GACjB,MAAM,YAAY,CAAC;AAGpB,OAAO,EAAE,eAAe,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAG7D,OAAO,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAGvF,OAAO,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AACtD,YAAY,EAAE,WAAW,EAAE,cAAc,EAAE,aAAa,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AACvG,OAAO,EAAE,iBAAiB,EAAE,MAAM,0BAA0B,CAAC;AAC7D,OAAO,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AACvD,OAAO,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AACvD,OAAO,EAAE,iBAAiB,EAAE,MAAM,2BAA2B,CAAC;AAG9D,OAAO,EAAE,YAAY,EAAE,oBAAoB,EAAE,sBAAsB,EAAE,MAAM,mBAAmB,CAAC;AAG/F,OAAO,EAAE,sBAAsB,EAAE,MAAM,8BAA8B,CAAC;AACtE,YAAY,EAAE,kBAAkB,EAAE,MAAM,8BAA8B,CAAC"}

package/dist/index.js

@@ -0,0 +1,35 @@
+/**
+ * Darwin — AI agents that improve themselves.
+ *
+ * @example
+ * ```typescript
+ * import { defineAgent, defineConfig, runAgent } from 'darwin-agents';
+ *
+ * const myAgent = defineAgent({
+ *   name: 'summarizer',
+ *   role: 'Text Summarizer',
+ *   systemPrompt: 'Summarize text in 3 bullet points.',
+ *   evolution: { enabled: true, evaluator: 'critic' },
+ * });
+ *
+ * const result = await runAgent(myAgent, 'Summarize this article...');
+ * ```
+ */
+// Core API
+export { defineAgent, defineConfig, loadConfig, loadConfigSync } from './core/agent.js';
+export { runAgent } from './core/runner.js';
+// Constants
+export { DEFAULT_WEIGHTS, DEFAULT_SAFETY } from './types.js';
+// Built-in Agents
+export { writer, researcher, critic, analyst, builtinAgents } from './agents/index.js';
+// Providers
+export { createProvider } from './providers/index.js';
+export { AnthropicProvider } from './providers/anthropic.js';
+export { OpenAIProvider } from './providers/openai.js';
+export { OllamaProvider } from './providers/ollama.js';
+export { ClaudeCliProvider } from './providers/claude-cli.js';
+// Memory
+export { createMemory, SqliteMemoryProvider, PostgresMemoryProvider } from './memory/index.js';
+// Notifications
+export { loadNotificationConfig } from './evolution/notifications.js';
+//# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.js","sourceRoot":"","sources":["../src/index.ts"],"names":[],"mappings":"AAAA;;;;;;;;;;;;;;;;GAgBG;AAEH,WAAW;AACX,OAAO,EAAE,WAAW,EAAE,YAAY,EAAE,UAAU,EAAE,cAAc,EAAE,MAAM,iBAAiB,CAAC;AACxF,OAAO,EAAE,QAAQ,EAAE,MAAM,kBAAkB,CAAC;AAsB5C,YAAY;AACZ,OAAO,EAAE,eAAe,EAAE,cAAc,EAAE,MAAM,YAAY,CAAC;AAE7D,kBAAkB;AAClB,OAAO,EAAE,MAAM,EAAE,UAAU,EAAE,MAAM,EAAE,OAAO,EAAE,aAAa,EAAE,MAAM,mBAAmB,CAAC;AAEvF,YAAY;AACZ,OAAO,EAAE,cAAc,EAAE,MAAM,sBAAsB,CAAC;AAEtD,OAAO,EAAE,iBAAiB,EAAE,MAAM,0BAA0B,CAAC;AAC7D,OAAO,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AACvD,OAAO,EAAE,cAAc,EAAE,MAAM,uBAAuB,CAAC;AACvD,OAAO,EAAE,iBAAiB,EAAE,MAAM,2BAA2B,CAAC;AAE9D,SAAS;AACT,OAAO,EAAE,YAAY,EAAE,oBAAoB,EAAE,sBAAsB,EAAE,MAAM,mBAAmB,CAAC;AAE/F,gBAAgB;AAChB,OAAO,EAAE,sBAAsB,EAAE,MAAM,8BAA8B,CAAC"}

package/dist/memory/index.d.ts

@@ -0,0 +1,32 @@
+/**
+ * Darwin — Memory Factory
+ *
+ * Creates the appropriate MemoryProvider based on config.
+ * Supports SQLite (free) and PostgreSQL (pro).
+ */
+import type { DarwinConfig, MemoryProvider } from '../types.js';
+/**
+ * Create a MemoryProvider based on the Darwin config.
+ *
+ * @param config - Darwin configuration with memory backend selection
+ * @returns An uninitialized MemoryProvider — call `init()` before use
+ *
+ * @example
+ * ```ts
+ * // SQLite (default, zero-config)
+ * const memory = createMemory({ provider: 'claude-cli', memory: 'sqlite' });
+ *
+ * // PostgreSQL (production, concurrent agents)
+ * const memory = createMemory({
+ *   provider: 'claude-cli',
+ *   memory: 'postgres',
+ *   postgresUrl: 'postgresql://user:pass@localhost:5432/darwin',
+ * });
+ *
+ * await memory.init();
+ * ```
+ */
+export declare function createMemory(config: DarwinConfig): MemoryProvider;
+export { SqliteMemoryProvider } from './sqlite-memory.js';
+export { PostgresMemoryProvider } from './postgres-memory.js';
+//# sourceMappingURL=index.d.ts.map

package/dist/memory/index.d.ts.map

@@ -0,0 +1 @@
+{"version":3,"file":"index.d.ts","sourceRoot":"","sources":["../../src/memory/index.ts"],"names":[],"mappings":"AAAA;;;;;GAKG;AAEH,OAAO,KAAK,EAAE,YAAY,EAAE,cAAc,EAAE,MAAM,aAAa,CAAC;AAIhE;;;;;;;;;;;;;;;;;;;;GAoBG;AACH,wBAAgB,YAAY,CAAC,MAAM,EAAE,YAAY,GAAG,cAAc,CAmBjE;AAED,OAAO,EAAE,oBAAoB,EAAE,MAAM,oBAAoB,CAAC;AAC1D,OAAO,EAAE,sBAAsB,EAAE,MAAM,sBAAsB,CAAC"}