agentic-qe 3.9.30 → 3.9.31

package/dist/learning/qe-reasoning-bank-types.d.ts

@@ -79,6 +79,17 @@ export interface QERoutingResult {
     guidance: string[];
     /** Reasoning for the recommendation */
     reasoning: string;
+    /**
+     * ADR-095 telemetry — all optional, populated by routeTask when the
+     * exploration / Q-value path runs. Callers persist these into
+     * routing_outcomes for retrospective analysis.
+     */
+    /** True if ε-greedy promoted an alternative over the greedy winner. */
+    exploration?: boolean;
+    /** Mincut safety multiplier in effect (1.0 = full rate, 0.2 = dampened). */
+    criticality?: number;
+    /** Average qWeight across the scored agents — proxy for Q-table maturity. */
+    qWeight?: number;
 }
 /**
  * Pattern learning outcome

package/dist/learning/qe-reasoning-bank.d.ts

@@ -118,6 +118,17 @@ export declare class QEReasoningBank implements IQEReasoningBank {
      * Route a task to optimal agent
      */
     routeTask(request: QERoutingRequest): Promise<Result<QERoutingResult>>;
+    /**
+     * ADR-095: Build a per-agent Q-value lookup closure for the current
+     * routing decision. The closure executes a prepared-statement query
+     * against `rl_q_values` for each agent considered.
+     *
+     * Returns a no-op lookup when the unified memory backend is unavailable
+     * (e.g. during early boot or in tests that don't initialize SQLite).
+     * The no-op causes `calculateAgentScores` to fall back to pure-static
+     * scoring — identical to pre-ADR-095 behavior.
+     */
+    private buildQValueLookup;
     /**
      * Get guidance for a domain
      */

package/dist/learning/qe-reasoning-bank.js

@@ -25,7 +25,9 @@ import { getWitnessChain } from '../audit/witness-chain.js';
 // Import extracted modules
 import { DEFAULT_QE_REASONING_BANK_CONFIG } from './qe-reasoning-bank-types.js';
 import { PRETRAINED_PATTERNS } from './pretrained-patterns.js';
-import { AGENT_CAPABILITIES, calculateAgentScores, } from './agent-routing.js';
+import { AGENT_CAPABILITIES, calculateAgentScores, applyExplorationPolicy, resolveExplorationRate, deriveTaskType, buildRoutingStateKey, deriveComplexityBucket, } from './agent-routing.js';
+import { resolveTopologyCriticalFromSharedMincut } from './routing-topology-gate.js';
+import { getUnifiedMemory } from '../kernel/unified-memory.js';
 import { resizeEmbedding, hashEmbedding } from './embedding-utils.js';
 import { checkPatternPromotionWithCoherence, promotePattern as promotePatternFn, seedCrossDomainPatterns as seedCrossDomainPatternsFn, } from './pattern-promotion.js';
 // ============================================================================
@@ -395,8 +397,35 @@ export class QEReasoningBank {
                     agentDomainPatternCounts.set(agentType, count);
                 }
             }
-            // 4. Calculate agent scores using extracted function
-            const agentScores = calculateAgentScores(detectedDomains, request.capabilities, agentDomainPatternCounts, this.config.routingWeights);
+            // ADR-095: build the Q-value lookup closure from the task's state_key.
+            // The state_key MUST match what `updateHookRouterQValue` writes from
+            // post-task — buildRoutingStateKey enforces that.
+            const stateKey = buildRoutingStateKey({
+                taskType: deriveTaskType(request.task),
+                priority: 'normal',
+                domain: detectedDomains[0],
+                complexityBucket: deriveComplexityBucket(request.task),
+            });
+            const qValueLookup = this.buildQValueLookup(stateKey);
+            // 4. Calculate agent scores (now Q-blended when lookup available)
+            const agentScores = calculateAgentScores(detectedDomains, request.capabilities, agentDomainPatternCounts, this.config.routingWeights, AGENT_CAPABILITIES, request.context?.language, qValueLookup);
+            // ADR-095: ε-greedy exploration with mincut safety gate.
+            // The exploration is best-effort — failures fall through to greedy.
+            // Topology check uses `resolveTopologyCriticalFromSharedMincut` —
+            // see that module for the empty-graph regression context.
+            let safetyMultiplier = 1.0;
+            try {
+                const topologyCritical = resolveTopologyCriticalFromSharedMincut();
+                const rateInfo = resolveExplorationRate({
+                    envOverride: process.env.AQE_ROUTER_EXPLORATION_RATE,
+                    topologyCritical,
+                });
+                safetyMultiplier = rateInfo.safetyMultiplier;
+                applyExplorationPolicy(agentScores, rateInfo.epsilon);
+            }
+            catch {
+                // Exploration failures are non-fatal; greedy order stands.
+            }
             const recommended = agentScores[0];
             const alternatives = agentScores.slice(1, 4);
             // Generate guidance
@@ -411,6 +440,11 @@ export class QEReasoningBank {
             }
             // Update stats
             this.stats.totalRoutingConfidence += recommended.score;
+            // ADR-095 telemetry: average qWeight across the scored agents — a
+            // proxy for Q-table maturity. Operators can correlate this with
+            // routing_outcomes.quality_score over time to see if learning helps.
+            const totalQWeight = agentScores.reduce((acc, a) => acc + (a.qWeight ?? 0), 0);
+            const avgQWeight = agentScores.length > 0 ? totalQWeight / agentScores.length : 0;
             const result = {
                 recommendedAgent: recommended.agent,
                 confidence: recommended.score,
@@ -419,6 +453,9 @@ export class QEReasoningBank {
                 patterns,
                 guidance,
                 reasoning: recommended.reasoning.join('; '),
+                exploration: recommended.exploration === true,
+                criticality: safetyMultiplier,
+                qWeight: avgQWeight,
             };
             return ok(result);
         }
@@ -426,6 +463,52 @@ export class QEReasoningBank {
             return err(toError(error));
         }
     }
+    /**
+     * ADR-095: Build a per-agent Q-value lookup closure for the current
+     * routing decision. The closure executes a prepared-statement query
+     * against `rl_q_values` for each agent considered.
+     *
+     * Returns a no-op lookup when the unified memory backend is unavailable
+     * (e.g. during early boot or in tests that don't initialize SQLite).
+     * The no-op causes `calculateAgentScores` to fall back to pure-static
+     * scoring — identical to pre-ADR-095 behavior.
+     */
+    buildQValueLookup(stateKey) {
+        try {
+            const um = getUnifiedMemory();
+            if (!um.isInitialized())
+                return () => undefined;
+            const db = um.getDatabase();
+            // Verify the Q-table exists; many CLI commands run without ever
+            // initializing the schema and the SELECT below would throw.
+            const tableExists = db
+                .prepare("SELECT 1 FROM sqlite_master WHERE type='table' AND name='rl_q_values'")
+                .get();
+            if (!tableExists)
+                return () => undefined;
+            const stmt = db.prepare(`
+        SELECT q_value, visits FROM rl_q_values
+         WHERE algorithm = 'q-learning'
+           AND agent_id = 'aqe-hook-router'
+           AND state_key = ?
+           AND action_key = ?
+      `);
+            return (agentType) => {
+                try {
+                    const row = stmt.get(stateKey, agentType);
+                    if (!row)
+                        return undefined;
+                    return { qValue: row.q_value, visits: row.visits };
+                }
+                catch {
+                    return undefined;
+                }
+            };
+        }
+        catch {
+            return () => undefined;
+        }
+    }
     /**
      * Get guidance for a domain
      */

package/dist/learning/routing-topology-gate.d.ts

@@ -0,0 +1,40 @@
+/**
+ * Routing Topology Safety Gate (ADR-095 — CRITICAL #1 follow-up)
+ *
+ * Resolves whether the agent dependency graph is in a "critical" topology
+ * state for the purpose of dampening exploration in QEReasoningBank.routeTask.
+ *
+ * Why this lives in its own module:
+ *
+ * 1. Testability. The reasoning-bank's full init path (real embeddings,
+ *    HNSW, pretrained patterns) is OOM-heavy in test environments. Putting
+ *    the topology gate here lets us unit-test it against a fresh
+ *    SwarmGraph singleton without spinning up the kernel + bank stack.
+ *
+ * 2. The empty-graph regression. `MinCutCalculator.getMinCutValue` returns
+ *    0 on an empty graph; `MinCutHealthMonitor.isCritical()` is defined as
+ *    `minCutValue < warningThreshold`, so an empty graph reports as
+ *    critical. CLI hook routing — the only place this gate fires in
+ *    practice — sees an empty graph because Queen coordinator activity
+ *    is what populates it. Without the emptiness check, exploration is
+ *    permanently dampened in the exact deployment case ADR-095 was meant
+ *    to help.
+ *
+ * Treat empty/uninitialized graph as "no signal" (full ε rate). Only
+ * consult `isCritical()` when the graph has actual vertices.
+ */
+/**
+ * Returns true ONLY when:
+ *   1. The shared monitor + graph are both initialized
+ *   2. The graph has at least one vertex (so the mincut value is real,
+ *      not the degenerate 0 of an empty graph)
+ *   3. The monitor's `isCritical()` returns true
+ *
+ * Any other state returns false (full ε rate / no dampening).
+ *
+ * All errors are swallowed — the gate is a best-effort signal that must
+ * never block routing. Caller defaults safetyMultiplier to 1.0 when this
+ * returns false.
+ */
+export declare function resolveTopologyCriticalFromSharedMincut(): boolean;
+//# sourceMappingURL=routing-topology-gate.d.ts.map

package/dist/learning/routing-topology-gate.js

@@ -0,0 +1,55 @@
+/**
+ * Routing Topology Safety Gate (ADR-095 — CRITICAL #1 follow-up)
+ *
+ * Resolves whether the agent dependency graph is in a "critical" topology
+ * state for the purpose of dampening exploration in QEReasoningBank.routeTask.
+ *
+ * Why this lives in its own module:
+ *
+ * 1. Testability. The reasoning-bank's full init path (real embeddings,
+ *    HNSW, pretrained patterns) is OOM-heavy in test environments. Putting
+ *    the topology gate here lets us unit-test it against a fresh
+ *    SwarmGraph singleton without spinning up the kernel + bank stack.
+ *
+ * 2. The empty-graph regression. `MinCutCalculator.getMinCutValue` returns
+ *    0 on an empty graph; `MinCutHealthMonitor.isCritical()` is defined as
+ *    `minCutValue < warningThreshold`, so an empty graph reports as
+ *    critical. CLI hook routing — the only place this gate fires in
+ *    practice — sees an empty graph because Queen coordinator activity
+ *    is what populates it. Without the emptiness check, exploration is
+ *    permanently dampened in the exact deployment case ADR-095 was meant
+ *    to help.
+ *
+ * Treat empty/uninitialized graph as "no signal" (full ε rate). Only
+ * consult `isCritical()` when the graph has actual vertices.
+ */
+import { getSharedMinCutGraph, getSharedMinCutMonitor, isSharedMinCutGraphInitialized, isSharedMinCutMonitorInitialized, } from '../coordination/mincut/shared-singleton.js';
+/**
+ * Returns true ONLY when:
+ *   1. The shared monitor + graph are both initialized
+ *   2. The graph has at least one vertex (so the mincut value is real,
+ *      not the degenerate 0 of an empty graph)
+ *   3. The monitor's `isCritical()` returns true
+ *
+ * Any other state returns false (full ε rate / no dampening).
+ *
+ * All errors are swallowed — the gate is a best-effort signal that must
+ * never block routing. Caller defaults safetyMultiplier to 1.0 when this
+ * returns false.
+ */
+export function resolveTopologyCriticalFromSharedMincut() {
+    try {
+        if (!isSharedMinCutMonitorInitialized())
+            return false;
+        if (!isSharedMinCutGraphInitialized())
+            return false;
+        const graph = getSharedMinCutGraph();
+        if (graph.isEmpty())
+            return false;
+        return getSharedMinCutMonitor().isCritical();
+    }
+    catch {
+        return false;
+    }
+}
+//# sourceMappingURL=routing-topology-gate.js.map

package/dist/learning/sqlite-persistence.d.ts

@@ -115,7 +115,12 @@ export declare class SQLitePatternStore {
         embedding: number[];
     }>;
     /**
-     * Record pattern usage
+     * Record pattern usage.
+     *
+     * Delegates to {@link recordPatternUsage} so the hook path
+     * (hooks-dream-learning.ts) and the canonical pattern-store path share a
+     * single writer — #486 Gap B. Preserves the previous throw-on-missing
+     * contract by checking the helper's `updated` flag.
      */
     recordUsage(patternId: string, success: boolean, metrics?: Record<string, unknown>, feedback?: string): void;
     /**

package/dist/learning/sqlite-persistence.js

@@ -17,6 +17,7 @@ import { safeJsonParse } from '../shared/safe-json.js';
 import { toErrorMessage } from '../shared/error-utils.js';
 import { getUnifiedMemory } from '../kernel/unified-memory.js';
 import { computeBatchEmbeddings, getEmbeddingDimension } from './real-embeddings.js';
+import { recordPatternUsage } from './pattern-usage-recorder.js';
 export const DEFAULT_SQLITE_CONFIG = {
     // LEGACY: Ignored when useUnified=true (the default). All data goes to memory.db
     dbPath: '.agentic-qe/memory.db',
@@ -518,32 +519,25 @@ export class SQLitePatternStore {
         }));
     }
     /**
-     * Record pattern usage
+     * Record pattern usage.
+     *
+     * Delegates to {@link recordPatternUsage} so the hook path
+     * (hooks-dream-learning.ts) and the canonical pattern-store path share a
+     * single writer — #486 Gap B. Preserves the previous throw-on-missing
+     * contract by checking the helper's `updated` flag.
      */
     recordUsage(patternId, success, metrics, feedback) {
         if (!this.db)
             throw new Error('Database not initialized');
-        const insertUsage = this.prepared.get('insertUsage');
-        const updatePattern = this.prepared.get('updatePattern');
-        if (!insertUsage || !updatePattern) {
-            throw new Error('Prepared statements not ready');
-        }
-        // Get current pattern for quality score calculation
-        const pattern = this.getPattern(patternId);
-        if (!pattern) {
+        const result = recordPatternUsage(this.db, {
+            patternId,
+            success,
+            metrics,
+            feedback,
+        });
+        if (!result.updated) {
             throw new Error(`Pattern not found: ${patternId}`);
         }
-        const newUsageCount = pattern.usageCount + 1;
-        const newSuccessfulUses = pattern.successfulUses + (success ? 1 : 0);
-        const newSuccessRate = newSuccessfulUses / newUsageCount;
-        // Quality score: confidence * 0.3 + usage * 0.2 + success_rate * 0.5
-        const usageScore = Math.min(1, newUsageCount / 100);
-        const qualityScore = pattern.confidence * 0.3 + usageScore * 0.2 + newSuccessRate * 0.5;
-        const transaction = this.db.transaction(() => {
-            insertUsage.run(patternId, success ? 1 : 0, metrics ? JSON.stringify(metrics) : null, feedback || null);
-            updatePattern.run(success ? 1 : 0, success ? 1 : 0, qualityScore, patternId);
-        });
-        transaction();
     }
     /**
      * Delete a pattern and its related data (embedding, usage history)