agentic-qe 3.10.0 → 3.10.2

package/dist/shared/utils/rabitq.d.ts

@@ -0,0 +1,133 @@
+/**
+ * RaBitQ-style 1-bit sign signatures for fast vector retrieval.
+ *
+ * Ports ruflo's perf-M4 optimization: sign-random-projection + Hamming popcount
+ * as a cheap prefilter ahead of exact cosine reranking.
+ *
+ * ## Why this works
+ *
+ * For unit-normalized embeddings, the sign of each coordinate is a valid
+ * Locality-Sensitive Hash (LSH) for cosine similarity: the probability that two
+ * vectors agree on the sign of a random projection is `1 - theta/pi`, where
+ * `theta` is the angle between them (Charikar 2002, "SimHash"). The AQE
+ * canonical embedding dimension is 384 (patterns.rvf dim=384); the coordinate
+ * basis itself acts as the random projection family, so `bit_i = (vec[i] >= 0)`
+ * yields a Hamming distance that is monotonically related to cosine angle.
+ *
+ * We therefore use Hamming distance over packed 1-bit signatures as a *cheap*
+ * ranking proxy to shrink the candidate pool, then run *exact* cosine on the
+ * survivors. With a large enough rerank pool the final top-k is identical to a
+ * full exact scan — only the work is reduced.
+ *
+ * ## Memory
+ *
+ * A 384-dim Float32 vector costs 1536 bytes. Its sign signature costs 48 bytes
+ * (12 × uint32). That is a 32× index-memory reduction for the prefilter stage.
+ *
+ * Self-contained: no I/O, no side effects, pure functions + one ranking helper.
+ *
+ * @module shared/utils/rabitq
+ */
+/** AQE canonical embedding dimension (patterns.rvf dim=384). */
+export declare const SIGNATURE_DIM = 384;
+/**
+ * Bytes per signature for the canonical 384-dim embedding:
+ * ceil(384 / 32) = 12 words × 4 bytes = 48 bytes.
+ * (vs 384 × 4 = 1536 bytes for the raw Float32 vector → 32× reduction.)
+ */
+export declare const SIGNATURE_BYTES: number;
+/**
+ * Build a 1-bit sign signature from a vector.
+ *
+ * `bit_i = (vec[i] >= 0)`. Bits are packed LSB-first into Uint32 words:
+ * coordinate `i` lives in word `i >> 5` at bit `i & 31`.
+ *
+ * For a 384-dim input this returns 12 words (48 bytes). Inputs of other
+ * dimensions are supported (signature length = ceil(dim / 32) words).
+ *
+ * @param vec Embedding vector (assumed unit-normalized for the LSH guarantee,
+ *            but the function works on any real vector).
+ * @returns Packed sign bits as a Uint32Array.
+ */
+export declare function signSignature(vec: number[] | Float32Array): Uint32Array;
+/**
+ * Hamming distance between two packed signatures: popcount(a XOR b) summed
+ * over all words.
+ *
+ * @throws If the two signatures have different word lengths.
+ */
+export declare function hammingDistance(a: Uint32Array, b: Uint32Array): number;
+/**
+ * Hamming similarity in [0, 1]: `1 - hammingDistance / bits`.
+ *
+ * This is a monotonic proxy for cosine similarity (higher = closer). `bits`
+ * should be the true embedding dimension (e.g. 384), not `words × 32`, so that
+ * trailing padding bits — which are always 0 in both signatures and never
+ * contribute to the distance — don't deflate the normalization.
+ */
+export declare function hammingSimilarity(a: Uint32Array, b: Uint32Array, bits: number): number;
+/** A retrieval candidate. `signature` is optional and lazily built if absent. */
+export interface RaBitQCandidate {
+    id: string;
+    vector: number[] | Float32Array;
+    /** Precomputed sign signature; built on demand from `vector` if omitted. */
+    signature?: Uint32Array;
+}
+/** A scored result. `score` is the exact cosine similarity (or Hamming proxy). */
+export interface RaBitQResult {
+    id: string;
+    score: number;
+}
+/** Tuning knobs for {@link topKBySignature}. */
+export interface TopKOptions {
+    /**
+     * Below this candidate count, skip RaBitQ entirely and do an exact cosine
+     * top-k. RaBitQ's signature-build + Hamming-scan overhead is not worth it for
+     * small N. Default 100 (matches ruflo's gate).
+     */
+    gateN?: number;
+    /**
+     * How many Hamming-ranked candidates to feed into the exact rerank.
+     *
+     * Larger pool → higher recall, more exact-cosine work. **Default is
+     * recall-aware**: `max(4 × k, ceil(0.4 × N))`.
+     *
+     * IMPORTANT — empirically measured tradeoff (384-dim, see
+     * scripts/benchmark-rabitq.ts): a 1-bit sign signature is a *coarse*
+     * quantization, so a small fixed pool (e.g. `4 × k = 40`) does NOT preserve
+     * recall@10 at scale — at N=1000 it captures only ~70-85% of the true top-10,
+     * and worse at N=5000. To keep recall@10 ≥ ~0.95 the pool must be a sizeable
+     * fraction of N (~40% empirically), which limits the achievable speedup.
+     * Callers that can tolerate lower recall may pass a smaller pool for more
+     * speed; callers that need exact output should not use RaBitQ at all.
+     */
+    rerankPool?: number;
+    /**
+     * If true (default), run full cosineSimilarity on the Hamming-top pool and
+     * return the true top-k by cosine. If false, return the Hamming-only ranking
+     * (faster, approximate).
+     */
+    exactRerank?: boolean;
+}
+/**
+ * Retrieve the top-k candidates for a query using a RaBitQ Hamming prefilter
+ * followed by an exact cosine rerank.
+ *
+ * Behavior:
+ * - `candidates.length < gateN` → plain exact cosine top-k (gate; RaBitQ
+ *   overhead not worth it for small N).
+ * - otherwise → build/reuse sign signatures, rank ALL candidates by Hamming
+ *   distance (cheap), take the top `rerankPool`, then:
+ *     - `exactRerank=true` (default): run FULL cosine on just that pool and
+ *       return the true top-k. Output equals an exact scan whenever the true
+ *       top-k all fall within the Hamming top-`rerankPool` — recall is
+ *       preserved by the rerank, only speed changes.
+ *     - `exactRerank=false`: return the Hamming-only top-k (approximate).
+ *
+ * @param query Query vector (same dimension as candidate vectors).
+ * @param candidates Candidate set.
+ * @param k Number of results to return.
+ * @param opts Tuning knobs (see {@link TopKOptions}).
+ */
+export declare function topKBySignature(query: number[] | Float32Array, candidates: RaBitQCandidate[], k: number, opts?: TopKOptions): RaBitQResult[];
+//# sourceMappingURL=rabitq.d.ts.map

package/dist/shared/utils/rabitq.js

@@ -0,0 +1,201 @@
+/**
+ * RaBitQ-style 1-bit sign signatures for fast vector retrieval.
+ *
+ * Ports ruflo's perf-M4 optimization: sign-random-projection + Hamming popcount
+ * as a cheap prefilter ahead of exact cosine reranking.
+ *
+ * ## Why this works
+ *
+ * For unit-normalized embeddings, the sign of each coordinate is a valid
+ * Locality-Sensitive Hash (LSH) for cosine similarity: the probability that two
+ * vectors agree on the sign of a random projection is `1 - theta/pi`, where
+ * `theta` is the angle between them (Charikar 2002, "SimHash"). The AQE
+ * canonical embedding dimension is 384 (patterns.rvf dim=384); the coordinate
+ * basis itself acts as the random projection family, so `bit_i = (vec[i] >= 0)`
+ * yields a Hamming distance that is monotonically related to cosine angle.
+ *
+ * We therefore use Hamming distance over packed 1-bit signatures as a *cheap*
+ * ranking proxy to shrink the candidate pool, then run *exact* cosine on the
+ * survivors. With a large enough rerank pool the final top-k is identical to a
+ * full exact scan — only the work is reduced.
+ *
+ * ## Memory
+ *
+ * A 384-dim Float32 vector costs 1536 bytes. Its sign signature costs 48 bytes
+ * (12 × uint32). That is a 32× index-memory reduction for the prefilter stage.
+ *
+ * Self-contained: no I/O, no side effects, pure functions + one ranking helper.
+ *
+ * @module shared/utils/rabitq
+ */
+import { cosineSimilarity } from './vector-math.js';
+// ============================================================================
+// Constants
+// ============================================================================
+/** Bits packed per Uint32 word. */
+const BITS_PER_WORD = 32;
+/** AQE canonical embedding dimension (patterns.rvf dim=384). */
+export const SIGNATURE_DIM = 384;
+/**
+ * Bytes per signature for the canonical 384-dim embedding:
+ * ceil(384 / 32) = 12 words × 4 bytes = 48 bytes.
+ * (vs 384 × 4 = 1536 bytes for the raw Float32 vector → 32× reduction.)
+ */
+export const SIGNATURE_BYTES = Math.ceil(SIGNATURE_DIM / BITS_PER_WORD) * 4; // 48
+// ============================================================================
+// Popcount (Brian-Kernighan)
+// ============================================================================
+/**
+ * Population count (number of set bits) of a 32-bit word.
+ *
+ * Uses the SWAR (SIMD-within-a-register) bit-twiddle which is branchless and
+ * faster than a per-byte table lookup for 32-bit words. The hdc-fingerprint
+ * module uses a 256-entry per-byte table + Brian-Kernighan for its Uint8Array
+ * vectors; here we operate on Uint32 words, so the SWAR variant is the natural
+ * fit and avoids importing/splitting words into bytes.
+ */
+function popcount32(n) {
+    // Force unsigned 32-bit semantics.
+    n = n >>> 0;
+    n = n - ((n >>> 1) & 0x55555555);
+    n = (n & 0x33333333) + ((n >>> 2) & 0x33333333);
+    n = (n + (n >>> 4)) & 0x0f0f0f0f;
+    return (Math.imul(n, 0x01010101) >>> 24) & 0xff;
+}
+// ============================================================================
+// Signature construction
+// ============================================================================
+/**
+ * Build a 1-bit sign signature from a vector.
+ *
+ * `bit_i = (vec[i] >= 0)`. Bits are packed LSB-first into Uint32 words:
+ * coordinate `i` lives in word `i >> 5` at bit `i & 31`.
+ *
+ * For a 384-dim input this returns 12 words (48 bytes). Inputs of other
+ * dimensions are supported (signature length = ceil(dim / 32) words).
+ *
+ * @param vec Embedding vector (assumed unit-normalized for the LSH guarantee,
+ *            but the function works on any real vector).
+ * @returns Packed sign bits as a Uint32Array.
+ */
+export function signSignature(vec) {
+    const dim = vec.length;
+    const words = Math.ceil(dim / BITS_PER_WORD);
+    const sig = new Uint32Array(words);
+    for (let i = 0; i < dim; i++) {
+        if (vec[i] >= 0) {
+            sig[i >> 5] |= 1 << (i & 31);
+        }
+    }
+    return sig;
+}
+// ============================================================================
+// Hamming distance / similarity
+// ============================================================================
+/**
+ * Hamming distance between two packed signatures: popcount(a XOR b) summed
+ * over all words.
+ *
+ * @throws If the two signatures have different word lengths.
+ */
+export function hammingDistance(a, b) {
+    if (a.length !== b.length) {
+        throw new Error(`Signature length mismatch: ${a.length} vs ${b.length} words`);
+    }
+    let dist = 0;
+    for (let i = 0; i < a.length; i++) {
+        dist += popcount32((a[i] ^ b[i]) >>> 0);
+    }
+    return dist;
+}
+/**
+ * Hamming similarity in [0, 1]: `1 - hammingDistance / bits`.
+ *
+ * This is a monotonic proxy for cosine similarity (higher = closer). `bits`
+ * should be the true embedding dimension (e.g. 384), not `words × 32`, so that
+ * trailing padding bits — which are always 0 in both signatures and never
+ * contribute to the distance — don't deflate the normalization.
+ */
+export function hammingSimilarity(a, b, bits) {
+    if (bits <= 0)
+        return 0;
+    return 1 - hammingDistance(a, b) / bits;
+}
+/**
+ * Exact cosine top-k over all candidates. Shared by the small-N gate and the
+ * rerank stage.
+ */
+function exactTopK(query, candidates, k) {
+    const scored = new Array(candidates.length);
+    for (let i = 0; i < candidates.length; i++) {
+        scored[i] = {
+            id: candidates[i].id,
+            score: cosineSimilarity(query, candidates[i].vector),
+        };
+    }
+    scored.sort((a, b) => b.score - a.score);
+    return scored.slice(0, k);
+}
+/**
+ * Retrieve the top-k candidates for a query using a RaBitQ Hamming prefilter
+ * followed by an exact cosine rerank.
+ *
+ * Behavior:
+ * - `candidates.length < gateN` → plain exact cosine top-k (gate; RaBitQ
+ *   overhead not worth it for small N).
+ * - otherwise → build/reuse sign signatures, rank ALL candidates by Hamming
+ *   distance (cheap), take the top `rerankPool`, then:
+ *     - `exactRerank=true` (default): run FULL cosine on just that pool and
+ *       return the true top-k. Output equals an exact scan whenever the true
+ *       top-k all fall within the Hamming top-`rerankPool` — recall is
+ *       preserved by the rerank, only speed changes.
+ *     - `exactRerank=false`: return the Hamming-only top-k (approximate).
+ *
+ * @param query Query vector (same dimension as candidate vectors).
+ * @param candidates Candidate set.
+ * @param k Number of results to return.
+ * @param opts Tuning knobs (see {@link TopKOptions}).
+ */
+export function topKBySignature(query, candidates, k, opts = {}) {
+    const gateN = opts.gateN ?? 100;
+    const exactRerank = opts.exactRerank ?? true;
+    const n = candidates.length;
+    if (n === 0 || k <= 0)
+        return [];
+    // Recall-aware default pool: a 1-bit signature in high dimensions is coarse,
+    // so the pool must scale with N to keep recall high (see rerankPool docs).
+    const rerankPool = opts.rerankPool ?? Math.max(4 * k, Math.ceil(0.4 * n));
+    // Small-N gate: exact cosine, skip signature machinery entirely.
+    if (n < gateN) {
+        return exactTopK(query, candidates, Math.min(k, n));
+    }
+    const bits = query.length;
+    const querySig = signSignature(query);
+    // Rank ALL candidates by Hamming distance (cheap popcount scan).
+    const hammingRanked = new Array(n);
+    for (let i = 0; i < n; i++) {
+        const cand = candidates[i];
+        const sig = cand.signature ?? signSignature(cand.vector);
+        hammingRanked[i] = { idx: i, dist: hammingDistance(querySig, sig) };
+    }
+    // Ascending distance = descending similarity.
+    hammingRanked.sort((a, b) => a.dist - b.dist);
+    const poolSize = Math.min(rerankPool, n);
+    if (!exactRerank) {
+        // Hamming-only ranking: return the top-k by Hamming proxy similarity.
+        const topK = Math.min(k, n);
+        const out = new Array(topK);
+        for (let i = 0; i < topK; i++) {
+            const { idx, dist } = hammingRanked[i];
+            out[i] = { id: candidates[idx].id, score: 1 - dist / bits };
+        }
+        return out;
+    }
+    // Exact rerank: full cosine over the Hamming-top pool, return true top-k.
+    const pool = new Array(poolSize);
+    for (let i = 0; i < poolSize; i++) {
+        pool[i] = candidates[hammingRanked[i].idx];
+    }
+    return exactTopK(query, pool, Math.min(k, poolSize));
+}
+//# sourceMappingURL=rabitq.js.map

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agentic-qe",
-  "version": "3.10.0",
+  "version": "3.10.2",
   "description": "Agentic Quality Engineering V3 - Domain-Driven Design Architecture with 13 Bounded Contexts, O(log n) coverage analysis, ReasoningBank learning, 60 specialized QE agents, mathematical Coherence verification, deep Claude Flow integration",
   "type": "module",
   "main": "./dist/index.js",
@@ -68,7 +68,8 @@
     "test:mcp:integration": "npm test -- --run tests/integration/mcp/",
     "test:integration:fast": "NODE_OPTIONS='--max-old-space-size=2048 --expose-gc' vitest run tests/integration/mcp/fleet-init-wires-daemon.test.ts tests/integration/workers/workers-reach-domains.test.ts tests/integration/learning/coordinator-roundtrip.test.ts tests/integration/bridge/bridge-end-to-end.test.ts",
     "test:integration": "NODE_OPTIONS='--max-old-space-size=4096 --expose-gc' vitest run tests/integration/ --exclude='**/browser/**' --exclude='**/browser-integration/**' --exclude='**/*.e2e.test.ts'",
-    "mcp:validate": "echo 'MCP validation: All tools registered in MCP server' && exit 0",
+    "mcp:validate": "node scripts/smoke-mcp-protocol.mjs",
+    "mcp:parity": "node scripts/audit-mcp-tool-parity.mjs",
     "mcp:report": "echo 'MCP Report: uses vitest for test reporting' && exit 0",
     "test:code-intelligence": "npm test -- --run tests/unit/domains/code-intelligence/ tests/unit/coordination/mincut/",
     "verify:counts": "node -e \"const fs=require('fs'),g=require('glob'),p=require('path');const src=g.globSync('src/**/*.ts',{ignore:'src/**/*.test.ts'});const tests=g.globSync('tests/**/*.test.ts');const agents=g.globSync('.claude/agents/v3/qe-*.md');const skills=g.globSync('.claude/skills/**/*.md');const ok=src.length>500&&tests.length>300;const r={status:ok?'pass':'fail',srcFiles:src.length,testFiles:tests.length,agents:agents.length,skills:skills.length};fs.mkdirSync('reports',{recursive:true});fs.writeFileSync(p.join('reports','verification-counts.json'),JSON.stringify(r,null,2));console.log('Counts:',JSON.stringify(r));if(!ok)process.exit(1)\"",
@@ -85,6 +86,10 @@
     "performance:gate": "npx tsx src/performance/run-gates.ts",
     "test:integration:browser": "VIBIUM_REAL_TESTS=true vitest run tests/integrations/vibium/vibium-real.test.ts",
     "benchmark:token-reduction": "vitest bench tests/benchmarks/code-intelligence-token-reduction.bench.ts",
+    "benchmark:self-learning": "tsx scripts/benchmark-self-learning.ts",
+    "benchmark:rabitq": "tsx scripts/benchmark-rabitq.ts",
+    "plugin:smoke": "bash plugins/agentic-qe-fleet/scripts/smoke.sh",
+    "pretrain:history": "tsx scripts/pretrain-from-history.ts",
     "dev": "node dist/cli/bundle.js",
     "start": "node dist/cli/bundle.js",
     "typecheck": "tsc --noEmit",
@@ -256,6 +261,7 @@
     ".claude/skills",
     ".claude/commands",
     ".claude/helpers",
+    ".claude/hooks",
     ".opencode/agents",
     ".opencode/skills",
     ".opencode/tools",

package/dist/cli/chunks/adapter-WTE6UVGP.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.10.0");process.exit(0)}
-import{a,b,c}from"./chunk-T7DLX3LS.js";import"./chunk-OGT45MZN.js";import"./chunk-TR7BZLB6.js";import"./chunk-OKGK7DBT.js";import"./chunk-KNL3QWVA.js";import"./chunk-55JPAF56.js";export{a as AgentBoosterAdapter,b as createAgentBoosterAdapter,c as createAgentBoosterAdapterSync};

package/dist/cli/chunks/aqe-learning-engine-LCLEBU7D.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.10.0");process.exit(0)}
-import{a,b,c,d,e,f}from"./chunk-OWMGD7FO.js";import"./chunk-5QJDH4Z5.js";import"./chunk-6H5MRVJS.js";import"./chunk-2GL4GH52.js";import"./chunk-C5QESAYA.js";import"./chunk-YHNEBCYQ.js";import"./chunk-WLLE54TA.js";import"./chunk-3OSCWD7Z.js";import"./chunk-F363JJUI.js";import"./chunk-VEOQH4W6.js";import"./chunk-LYVFC7C7.js";import"./chunk-U56TIYGP.js";import"./chunk-KCHFF4IE.js";import"./chunk-3BO7EKGO.js";import"./chunk-UE3XXKLN.js";import"./chunk-TEJPHJMW.js";import"./chunk-OKGK7DBT.js";import"./chunk-DXV6NRG3.js";import"./chunk-X73CRYF4.js";import"./chunk-BQT4J3BD.js";import"./chunk-FEBXP74Y.js";import"./chunk-KNL3QWVA.js";import"./chunk-MJBXQXSX.js";import"./chunk-ZCNVFULO.js";import"./chunk-QU54GUEA.js";import"./chunk-O3NAUNFC.js";import"./chunk-I4E6CLC4.js";import"./chunk-NT4PI5HI.js";import"./chunk-66GIKUI2.js";import"./chunk-SGONA5GS.js";import"./chunk-5UOV7T36.js";import"./chunk-KOSKGZK4.js";import"./chunk-55JPAF56.js";export{b as AQELearningEngine,a as DEFAULT_ENGINE_CONFIG,c as DEFAULT_META_LEARNING_CONFIG,d as MetaLearningEngine,e as createAQELearningEngine,f as createDefaultLearningEngine};

package/dist/cli/chunks/base-73I73HBF.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.10.0");process.exit(0)}
-import{a,b,c,d,e,f}from"./chunk-6MONUYQ5.js";import"./chunk-LQ3TA22E.js";import"./chunk-UE3XXKLN.js";import"./chunk-TEJPHJMW.js";import"./chunk-OKGK7DBT.js";import"./chunk-X73CRYF4.js";import"./chunk-BQT4J3BD.js";import"./chunk-KNL3QWVA.js";import"./chunk-MJBXQXSX.js";import"./chunk-ZCNVFULO.js";import"./chunk-QU54GUEA.js";import"./chunk-O3NAUNFC.js";import"./chunk-I4E6CLC4.js";import"./chunk-NT4PI5HI.js";import"./chunk-66GIKUI2.js";import"./chunk-SGONA5GS.js";import"./chunk-5UOV7T36.js";import"./chunk-KOSKGZK4.js";import"./chunk-55JPAF56.js";export{f as MCPToolBase,e as defaultToolLogger,d as getMemoryBackend,b as getSharedMemoryBackend,a as registerRvfResetFn,c as resetSharedMemoryBackend};

package/dist/cli/chunks/browser-workflow-2NSV5O6W.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.10.0");process.exit(0)}
-import{a}from"./chunk-6F3WJOU2.js";import"./chunk-6MONUYQ5.js";import"./chunk-LQ3TA22E.js";import"./chunk-UE3XXKLN.js";import"./chunk-TEJPHJMW.js";import"./chunk-OKGK7DBT.js";import"./chunk-X73CRYF4.js";import"./chunk-BQT4J3BD.js";import"./chunk-KNL3QWVA.js";import"./chunk-MJBXQXSX.js";import"./chunk-ZCNVFULO.js";import"./chunk-QU54GUEA.js";import"./chunk-O3NAUNFC.js";import"./chunk-I4E6CLC4.js";import"./chunk-NT4PI5HI.js";import"./chunk-66GIKUI2.js";import"./chunk-SGONA5GS.js";import"./chunk-5UOV7T36.js";import"./chunk-KOSKGZK4.js";import"./chunk-55JPAF56.js";export{a as BrowserWorkflowTool};

package/dist/cli/chunks/chunk-6MONUYQ5.js

@@ -1,2 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.10.0");process.exit(0)}
-import{a as M}from"./chunk-LQ3TA22E.js";import{f as y,o as T}from"./chunk-UE3XXKLN.js";import{a as p,c as S}from"./chunk-KNL3QWVA.js";import{i as b}from"./chunk-MJBXQXSX.js";T();S();import*as c from"path";import*as d from"fs";var m=null,u=null,l=null;function R(o){l=o}async function P(){return m||u||(u=(async()=>{let o=y(),e=c.join(o,".agentic-qe");d.existsSync(e)||d.mkdirSync(e,{recursive:!0});let r=new M({sqlite:{path:c.join(e,"memory.db"),walMode:!0,poolSize:3,busyTimeout:5e3},enableFallback:!0,defaultNamespace:"mcp-tools"});return await r.initialize(),m=r,r})(),u)}function $(){if(l)try{l()}catch{}m&&(m.dispose().catch(()=>{}),m=null),u=null}async function D(o){return o?.memory?o.memory:P()}var C={info:(o,e)=>console.log(`[MCP-TOOL] ${o}`,e||""),warn:(o,e)=>console.warn(`[MCP-TOOL] \u26A0\uFE0F ${o}`,e||""),error:(o,e)=>console.error(`[MCP-TOOL] \u274C ${o}`,e||"")},h=class{logger=C;currentDataSource="real";setLogger(e){this.logger=e}markAsDemoData(e,r){this.currentDataSource=e.demoMode?"demo":"fallback",this.logger.warn(`${this.config.name} returning ${this.currentDataSource} data`,{reason:r,requestId:e.requestId,demoMode:e.demoMode})}markAsRealData(){this.currentDataSource="real"}isDemoMode(e){return e.demoMode===!0}validate(e){let r=[];if(typeof e!="object"||e===null)return{valid:!1,errors:["Parameters must be an object"]};let t=e,{properties:a,required:n}=this.config.schema;if(n)for(let i of n)i in t||r.push(`Missing required field: ${i}`);for(let[i,s]of Object.entries(t)){let g=a[i];if(!g)continue;let f=this.validateType(i,s,g);f&&r.push(f)}return{valid:r.length===0,errors:r}}validateType(e,r,t){if(r==null)return null;let a=Array.isArray(r)?"array":typeof r;if(t.type==="array"&&!Array.isArray(r))return`${e} must be an array`;if(t.type!=="array"&&a!==t.type)return`${e} must be of type ${t.type}, got ${a}`;if(t.enum&&!t.enum.includes(r))return`${e} must be one of: ${t.enum.join(", ")}`;if(t.type==="number"){let n=r;if(t.minimum!==void 0&&n<t.minimum)return`${e} must be >= ${t.minimum}`;if(t.maximum!==void 0&&n>t.maximum)return`${e} must be <= ${t.maximum}`}if(t.type==="string"){let n=r;if(t.minLength!==void 0&&n.length<t.minLength)return`${e} must be at least ${t.minLength} characters`;if(t.maxLength!==void 0&&n.length>t.maxLength)return`${e} must be at most ${t.maxLength} characters`}return null}async invoke(e,r={}){let t=Date.now(),a=b();this.currentDataSource="real";let n=this.validate(e);if(!n.valid)return{success:!1,error:`Validation failed: ${n.errors.join("; ")}`,metadata:this.createMetadata(t,a)};let i={requestId:a,startTime:t,streaming:r.streaming,onStream:r.onStream,abortSignal:r.abortSignal,demoMode:r.demoMode};try{let s=await this.execute(e,i);return{...s,metadata:{...s.metadata,...this.createMetadata(t,a)}}}catch(s){return{success:!1,error:p(s),metadata:this.createMetadata(t,a)}}}createMetadata(e,r){return{executionTime:Date.now()-e,timestamp:new Date().toISOString(),requestId:r,domain:this.config.domain,toolName:this.config.name,dataSource:this.currentDataSource}}emitStream(e,r){e.streaming&&e.onStream&&e.onStream(r)}isAborted(e){return e.abortSignal?.aborted??!1}get name(){return this.config.name}get description(){return this.config.description}get domain(){return this.config.domain}getSchema(){return this.config.schema}get timeout(){return this.config.timeout??6e4}get supportsStreaming(){return this.config.streaming??!1}resetInstanceCache(){}};export{R as a,P as b,$ as c,D as d,C as e,h as f};

package/dist/cli/chunks/chunk-QL3U5VSM.js

@@ -1,62 +0,0 @@
-import{createRequire as __cr}from"module";const require=__cr(import.meta.url);if(process.argv.includes('--version')||process.argv.includes('-v')){console.log("3.10.0");process.exit(0)}
-import{p as D,q as z}from"./chunk-J5RJYFRM.js";import{k as A,o as R}from"./chunk-UE3XXKLN.js";import{i as g}from"./chunk-MJBXQXSX.js";import{b as w,c as T}from"./chunk-SGONA5GS.js";import{g as S}from"./chunk-55JPAF56.js";R();var b=null;function I(){if(b)return b;let m=(z(),S(D));if(!m.HierarchicalNSW)throw new Error("hnswlib-node module missing HierarchicalNSW export");return b=m.HierarchicalNSW,b}function M(){return!0}var E=class{indexes;config;initialized;nextId;unifiedAdapters=null;useUnified;constructor(e={}){this.config={M:e.M||16,efConstruction:e.efConstruction||200,efSearch:e.efSearch||50,dimension:e.dimension||384,metric:e.metric||"cosine",quantization:e.quantization||"none"},this.indexes=new Map,this.initialized=new Set,this.nextId=new Map,this.useUnified=M(),this.useUnified&&(this.unifiedAdapters=new Map)}initializeIndex(e){if(this.initialized.has(e))return;if(this.useUnified&&this.unifiedAdapters)try{let{HnswAdapter:s}=(T(),S(w)),o=new s(`embedding-${e}`,{dimensions:this.config.dimension,M:this.config.M,efConstruction:this.config.efConstruction,efSearch:this.config.efSearch,metric:this.config.metric==="dotproduct"?"cosine":this.config.metric});this.unifiedAdapters.set(e,o),this.initialized.add(e),this.nextId.set(e,0);return}catch{}let t={cosine:"cosine",euclidean:"l2",dotproduct:"ip"}[this.config.metric]||"cosine",n=I(),r=new n(t,this.config.dimension);r.initIndex({maxElements:1e4,m:this.config.M,efConstruction:this.config.efConstruction}),this.indexes.set(e,r),this.initialized.add(e),this.nextId.set(e,0)}addEmbedding(e,i){let t=e.namespace;this.initialized.has(t)||this.initializeIndex(t);let n=i??this.nextId.get(t);i===void 0&&this.nextId.set(t,n+1);let r=this.unifiedAdapters?.get(t);if(r){let d=e.vector instanceof Float32Array?e.vector:new Float32Array(this.toFloatArray(e.vector));return r.add(n,d),n}let s=this.indexes.get(t),o=this.toFloatArray(e.vector);return s.addPoint(o,n),n}addEmbeddingsBatch(e){return e.map(i=>this.addEmbedding(i.embedding,i.id))}search(e,i={}){let t=i.namespace||e.namespace;if(!this.initialized.has(t))return[];let n=i.limit||10,r=this.unifiedAdapters?.get(t);if(r){let l=e.vector instanceof Float32Array?e.vector:new Float32Array(this.toFloatArray(e.vector));return r.search(l,n).map(a=>({id:a.id,distance:1-a.score}))}let s=this.indexes.get(t),o=this.toFloatArray(e.vector),d=s.searchKnn(o,n);return d.neighbors.map((l,h)=>({id:l,distance:d.distances[h]}))}getIndexStats(e){if(!this.initialized.has(e))return null;let i=this.unifiedAdapters?.get(e);return i?{size:i.size(),maxElements:1e4,dimension:i.dimensions(),metric:this.config.metric}:{size:this.indexes.get(e).getCurrentCount(),maxElements:1e4,dimension:this.config.dimension,metric:this.config.metric}}async saveIndex(e,i){if(!this.initialized.has(e))throw new Error(`Namespace ${e} not initialized`);if(this.unifiedAdapters?.has(e)){console.warn(`[HNSWEmbeddingIndex] saveIndex() is a no-op for namespace '${e}' \u2014 unified HnswAdapter manages persistence internally.`);return}await this.indexes.get(e).writeIndex(i)}async loadIndex(e,i){if(this.useUnified){console.warn(`[HNSWEmbeddingIndex] loadIndex() is a no-op for namespace '${e}' \u2014 unified HnswAdapter manages persistence internally. Initialize via initializeIndex() instead.`);return}let n={cosine:"cosine",euclidean:"l2",dotproduct:"ip"}[this.config.metric]||"cosine",r=I(),s=new r(n,this.config.dimension);await s.readIndex(i),this.indexes.set(e,s),this.initialized.add(e)}clearIndex(e){if(this.initialized.has(e)){let i=this.unifiedAdapters?.get(e);i&&(i.clear?.(),this.unifiedAdapters.delete(e)),this.indexes.delete(e),this.initialized.delete(e),this.nextId.delete(e)}}clearAll(){if(this.unifiedAdapters){for(let e of this.unifiedAdapters.values())e.clear?.();this.unifiedAdapters.clear()}this.indexes.clear(),this.initialized.clear(),this.nextId.clear()}resizeIndex(e,i){this.initialized.has(e)&&(this.clearIndex(e),this.initializeIndex(e))}setEfSearch(e){this.config.efSearch=e}toFloatArray(e){if(Array.isArray(e))return e;if(e instanceof Float32Array)return Array.from(e);if(e instanceof Int8Array){let i=new Array(e.length);for(let t=0;t<e.length;t++)i[t]=e[t]/128;return i}if(e instanceof Uint8Array){let i=new Array(e.length);for(let t=0;t<e.length;t++)i[t]=(e[t]-128)/128;return i}throw new Error(`Unsupported vector type: ${typeof e}`)}getConfig(){return{...this.config}}isInitialized(e){return this.initialized.has(e)}getInitializedNamespaces(){return Array.from(this.initialized)}getSize(e){if(!this.initialized.has(e))return 0;let i=this.unifiedAdapters?.get(e);return i?i.size():this.nextId.get(e)||0}};var H={mergeSimilarityThreshold:.85,maxMergesPerRun:50,softThreshold:400,hardThreshold:2e3,archiveMinAgeDays:30,archiveQualityThreshold:.15,mergeQualityBoost:.02},N=class{config;db=null;initialized=!1;constructor(e={}){this.config={...H,...e}}async initialize(e){if(!this.initialized){if(e)this.db=e;else{let i=A();await i.initialize(),this.db=i.getDatabase()}this.initialized=!0}}async consolidateAll(e){this.ensureInitialized();let i={merged:0,qualityUpdated:0,archived:0,hardDeleted:0,activeRemaining:0,domainsProcessed:[]},t=this.db.prepare("SELECT domain, COUNT(*) as cnt FROM captured_experiences WHERE consolidated_into IS NULL GROUP BY domain").all(),n=e?t.filter(s=>e.includes(s.domain)):t.filter(s=>s.cnt>this.config.softThreshold);for(let{domain:s,cnt:o}of n){let d=await this.consolidateDomain(s,o);i.merged+=d.merged,i.qualityUpdated+=d.qualityUpdated,i.archived+=d.archived,i.hardDeleted+=d.hardDeleted,i.domainsProcessed.push(s)}let r=this.db.prepare("SELECT COUNT(*) as cnt FROM captured_experiences WHERE consolidated_into IS NULL").get();return i.activeRemaining=r.cnt,(i.merged>0||i.archived>0)&&console.log(`[ExperienceConsolidator] Consolidated: ${i.merged} merged, ${i.archived} archived, ${i.activeRemaining} active`),i}async consolidateDomain(e,i){this.ensureInitialized();let t={merged:0,qualityUpdated:0,archived:0,hardDeleted:0,activeRemaining:0,domainsProcessed:[e]};t.merged=await this.clusterAndMerge(e),t.qualityUpdated=this.reinforceQuality(e),t.archived=this.archiveValueless(e);let n=this.db.prepare("SELECT COUNT(*) as cnt FROM captured_experiences WHERE domain = ? AND consolidated_into IS NULL").get(e).cnt,r=0;return n>this.config.hardThreshold&&(r=this.hardDeleteExcess(e,n),t.archived+=r),t.activeRemaining=n-r,t}async bootstrapDomain(e){if(this.ensureInitialized(),this.db.prepare("SELECT value FROM kv_store WHERE key = ?").get(`consolidation_bootstrap_${e}`))return console.log(`[ExperienceConsolidator] Domain ${e} already bootstrapped`),{merged:0,qualityUpdated:0,archived:0,hardDeleted:0,activeRemaining:0,domainsProcessed:[e]};let t={...this.config};this.config.mergeSimilarityThreshold=.8,this.config.maxMergesPerRun=200;let n=await this.consolidateDomain(e);Object.assign(this.config,t);try{this.db.prepare("INSERT OR REPLACE INTO kv_store (key, value, namespace) VALUES (?, ?, 'system')").run(`consolidation_bootstrap_${e}`,new Date().toISOString())}catch{}return console.log(`[ExperienceConsolidator] Bootstrap ${e}: ${n.merged} merged, ${n.archived} archived`),n}async clusterAndMerge(e){let i=this.db.prepare(`
-      SELECT id, task, domain, quality, success, application_count,
-             consolidation_count, reuse_success_count, reuse_failure_count,
-             embedding, embedding_dimension, started_at
-      FROM captured_experiences
-      WHERE domain = ? AND consolidated_into IS NULL AND embedding IS NOT NULL
-      ORDER BY quality DESC
-    `).all(e);if(i.length<2)return 0;let t=new E({dimension:384,M:16,efConstruction:200,efSearch:100,metric:"cosine"}),n=new Map,r=0;for(let a of i)if(a.embedding&&a.embedding_dimension){let x=this.bufferToFloatArray(a.embedding,a.embedding_dimension),p=r++;n.set(p,a.id);let y={vector:x,dimension:384,namespace:"experiences",text:a.id,timestamp:Date.now(),quantization:"none",metadata:{}};t.addEmbedding(y,p)}let s=new Set,o=0,d=this.db.prepare("UPDATE captured_experiences SET consolidated_into = ? WHERE id = ?"),l=this.db.prepare(`
-      UPDATE captured_experiences
-      SET consolidation_count = consolidation_count + ?,
-          quality = MIN(1.0, quality + ?),
-          quality_updated_at = datetime('now')
-      WHERE id = ?
-    `),h=this.db.prepare(`
-      INSERT INTO experience_consolidation_log (id, domain, action, source_ids, target_id, details, created_at)
-      VALUES (?, ?, ?, ?, ?, ?, datetime('now'))
-    `);for(let a of i){if(o>=this.config.maxMergesPerRun)break;if(s.has(a.id)||!a.embedding||!a.embedding_dimension)continue;let p={vector:this.bufferToFloatArray(a.embedding,a.embedding_dimension),dimension:384,namespace:"experiences",text:a.id,timestamp:Date.now(),quantization:"none",metadata:{}},y=t.search(p,{limit:6,namespace:"experiences"}),u=[];for(let{id:v,distance:f}of y){if(o>=this.config.maxMergesPerRun)break;let c=n.get(v);if(!c||c===a.id||s.has(c)||1-f<this.config.mergeSimilarityThreshold)continue;let _=i.find(C=>C.id===c);_&&_.quality<=a.quality&&_.application_count===0&&(s.add(c),u.push(c),o++)}u.length>0&&this.db.transaction(()=>{for(let c of u)d.run(a.id,c);let f=this.config.mergeQualityBoost*u.length;l.run(u.length,f,a.id),h.run(g(),e,"merge",JSON.stringify(u),a.id,JSON.stringify({count:u.length,boost:f}))})()}return t.clearIndex("experiences"),o}reinforceQuality(e){let i=this.db.prepare(`
-      SELECT ce.id, ce.quality, ce.consolidation_count,
-             ce.reuse_success_count, ce.reuse_failure_count,
-             COALESCE(app.total, 0) as app_total,
-             COALESCE(app.successes, 0) as app_successes
-      FROM captured_experiences ce
-      LEFT JOIN (
-        SELECT experience_id,
-               COUNT(*) as total,
-               SUM(CASE WHEN success = 1 THEN 1 ELSE 0 END) as successes
-        FROM experience_applications
-        GROUP BY experience_id
-      ) app ON app.experience_id = ce.id
-      WHERE ce.domain = ? AND ce.consolidated_into IS NULL
-        AND (app.total > 0 OR ce.reuse_success_count > 0 OR ce.reuse_failure_count > 0)
-    `).all(e);if(i.length===0)return 0;let t=this.db.prepare(`
-      UPDATE captured_experiences
-      SET quality = ?, quality_updated_at = datetime('now')
-      WHERE id = ?
-    `),n=0;return this.db.transaction(()=>{for(let s of i){let o=s.app_total+s.reuse_success_count+s.reuse_failure_count,d=s.app_successes+s.reuse_success_count,l=o>0?d/o:0,h=Math.min(s.consolidation_count,10)/10,a=Math.min(1,.4*s.quality+.4*l+.2*h);Math.abs(a-s.quality)>.01&&(t.run(a,s.id),n++)}})(),n}archiveValueless(e){let i=new Date;i.setDate(i.getDate()-this.config.archiveMinAgeDays);let t=i.toISOString().replace("T"," ").slice(0,19),n=this.db.prepare(`
-      UPDATE captured_experiences
-      SET consolidated_into = 'archived'
-      WHERE domain = ?
-        AND consolidated_into IS NULL
-        AND quality < ?
-        AND application_count = 0
-        AND success = 0
-        AND consolidation_count = 1
-        AND started_at < ?
-    `).run(e,this.config.archiveQualityThreshold,t);if(n.changes>0)try{this.db.prepare(`
-          INSERT INTO experience_consolidation_log (id, domain, action, source_ids, details, created_at)
-          VALUES (?, ?, 'archive', '[]', ?, datetime('now'))
-        `).run(g(),e,JSON.stringify({count:n.changes}))}catch{}return n.changes}hardDeleteExcess(e,i){let t=i-this.config.hardThreshold;if(t<=0)return 0;let r=this.db.prepare(`
-      UPDATE captured_experiences
-      SET consolidated_into = 'archived'
-      WHERE id IN (
-        SELECT id FROM captured_experiences
-        WHERE domain = ? AND consolidated_into IS NULL AND application_count = 0
-        ORDER BY quality ASC, started_at ASC
-        LIMIT ?
-      )
-    `).run(e,t).changes;if(r>0){try{this.db.prepare(`
-          INSERT INTO experience_consolidation_log (id, domain, action, source_ids, details, created_at)
-          VALUES (?, ?, 'safety-valve-archive', '[]', ?, datetime('now'))
-        `).run(g(),e,JSON.stringify({count:r,currentCount:i,hardThreshold:this.config.hardThreshold,note:"soft-archive replacement for legacy hard-delete safety valve"}))}catch{}console.warn(`[ExperienceConsolidator] Safety valve: soft-archived ${r} from ${e} (was ${i}, threshold ${this.config.hardThreshold})`)}return r}ensureInitialized(){if(!this.initialized||!this.db)throw new Error("ExperienceConsolidator not initialized. Call initialize() first.")}bufferToFloatArray(e,i){let t=[];for(let n=0;n<i;n++)t.push(e.readFloatLE(n*4));return t}};export{E as a,N as b};