agentic-qe 3.7.14 → 3.7.16

package/.claude/helpers/brain-checkpoint.cjs

@@ -38,6 +38,17 @@ function exportBrain() {
   }
 
   try {
+    // Checkpoint WAL to ensure all pending writes are flushed to main DB
+    try {
+      execSync(`sqlite3 "${DB_PATH}" "PRAGMA wal_checkpoint(TRUNCATE);"`, {
+        timeout: 10000,
+        encoding: 'utf-8',
+      });
+      log('WAL checkpoint completed');
+    } catch (walErr) {
+      log(`WAL checkpoint warning: ${walErr.message}`);
+    }
+
     // Remove existing to avoid LockHeld errors
     if (fs.existsSync(RVF_PATH)) {
       fs.unlinkSync(RVF_PATH);

package/.claude/skills/skills-manifest.json

@@ -904,7 +904,7 @@
   },
   "metadata": {
     "generatedBy": "Agentic QE Fleet",
-    "fleetVersion": "3.7.14",
+    "fleetVersion": "3.7.16",
     "manifestVersion": "1.3.0",
     "lastUpdated": "2026-02-04T00:00:00.000Z",
     "contributors": [

package/CHANGELOG.md

@@ -5,6 +5,55 @@ All notable changes to the Agentic QE project will be documented in this file.
 The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.0.0/),
 and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
 
+## [3.7.16] - 2026-03-10
+
+### Added
+
+- **Tier 3 baseline collection and instrumentation** — Collect benchmark baselines for all Tier 3 features and add Priority 2 instrumentation: routing tier tags, HNSW/FTS5 search latency tracking, token tracker auto-save, and pipeline step timers.
+- **MCP persistence pipeline fix** — Wire `recordDomainFeedback` into the feedback loop (handler-factory Step 5d) so `test_outcomes` and `coverage_sessions` receive data from live MCP tool calls.
+- **Quality feedback loop singleton** — `getQualityFeedbackLoop()` provides cross-module access to the feedback loop instance.
+- **Experience embedding on capture** — Embedding computation now runs automatically when experiences are captured.
+- **Routing tier tracking** — New `model_tier` column in `routing_outcomes` with tier inference for cost analysis.
+- **Search latency instrumentation** — `performance.now()` timing added to HNSW search and FTS5 `searchFTS` for benchmarking.
+- **Token metrics auto-persistence** — `TokenMetricsCollector` now saves to DB automatically on initialization.
+- **Pipeline step latencies** — `TestSchedulingPipeline` results now include per-step timing data.
+
+### Fixed
+
+- **Critical: Test DB isolation** — Fixed `goap-planner.test.ts` and `q-value-persistence.test.ts` using relative `.agentic-qe/memory.db` paths that deleted the production database during test cleanup. Tests now use `os.tmpdir()`.
+- **Critical: Project root cache leak** — `resetUnifiedMemory()` now calls `clearProjectRootCache()` to prevent stale path cache from redirecting tests to the production DB.
+- **DB path safety redirect** — `UnifiedMemoryManager._doInitialize()` now detects and redirects when a test process (with `AQE_PROJECT_ROOT` set) tries to open a production `.agentic-qe/memory.db`.
+- **`process.cwd()` DB path bypasses** — `pull-agent.ts` and `brain-handler.ts` now use `findProjectRoot()` instead of `process.cwd()` to resolve the DB path, respecting `AQE_PROJECT_ROOT`.
+- **Optional native module graceful degradation** — FlashAttention and DecisionTransformer now degrade gracefully when native modules are unavailable.
+- **WAL checkpoint before RVF export** — Brain checkpoint now runs a WAL checkpoint before RVF export to ensure data consistency.
+
+## [3.7.15] - 2026-03-09
+
+### Added
+
+- **Proof-of-Quality CLI command** — `aqe prove` generates verifiable quality attestations with SHA-256 hashes, proving that tests, coverage, and security checks were actually run.
+- **OPD remediation hints** — Outcome-Pattern-Deviation engine now explains *why* a pattern failed and *how* to fix it ("bad because X, fix by Y").
+- **Per-agent MCP tool scoping** — Agents can now be restricted to specific MCP tool subsets for security isolation and least-privilege enforcement.
+- **Daily Markdown learning log** — Automatic audit trail of all learning activity per day, with session summaries and pattern metrics.
+- **FTS5 hybrid search** — Full-text search virtual table combining vector similarity with text matching for more accurate pattern retrieval.
+- **Binary reward assignment** — Outcome-based learning now uses clean +1/-1/0 rewards for sharper signal.
+- **Temporal decay on search** — Search results now apply a 30-day half-life decay, surfacing recent patterns over stale ones.
+- **Pre-compaction flush hook** — Prevents knowledge loss by flushing pending writes before database compaction.
+- **Embedding backfill script** — `scripts/backfill-embeddings.mjs` fills in missing embeddings for legacy patterns.
+
+### Fixed
+
+- **Promotion threshold inconsistency** — Standardized pattern promotion threshold to 3 across all code paths.
+- **Queen governance test flakiness** — Fixed singleton contamination in domain-handlers, handler-factory, and e2e test suites.
+- **MCP server domain tool failures** — Resolved timeout and invocation failures in domain-scoped MCP tool tests.
+- **Devil's advocate scoring** — Fixed FTS normalization, score capping, decay, and freshness calculations.
+
+### Changed
+
+- Coherence gate enabled by default for all learning sessions.
+- Configurable promotion activity window replaces hardcoded 7-day window.
+- 18,589 tests passing across 610 test files.
+
 ## [3.7.14] - 2026-03-08
 
 ### Added