agentic-qe 3.8.10 → 3.8.12

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

@@ -932,7 +932,7 @@
   },
   "metadata": {
     "generatedBy": "Agentic QE Fleet",
-    "fleetVersion": "3.8.10",
+    "fleetVersion": "3.8.12",
     "manifestVersion": "1.3.0",
     "lastUpdated": "2026-02-04T00:00:00.000Z",
     "contributors": [

package/CHANGELOG.md

@@ -5,6 +5,46 @@ 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).
 
+## [3.8.12] - 2026-03-29
+
+### Added
+
+- **RuVector Phase 5 — Pattern Intelligence (ADR-087)** — HDC pattern fingerprinting with 10K-bit binary hypervectors for O(1) XOR-based similarity, CUSUM drift detection across all 4 coherence gate types, and delta event sourcing with rollback via SQLite. EWC++ three-loop activation wired into quality-assessment, test-generation, and contract-testing coordinators.
+- **RuVector Phase 5 — Graph Learning (ADR-087)** — GraphMAE self-supervised learning with SPSA optimizer for masked graph autoencoders, Modern Hopfield networks for exponential-capacity exact pattern recall, and cold-tier GNN training with LRU-cached mini-batch trainer and file-backed larger-than-RAM graphs.
+- **RuVector Phase 5 — Scale, Optimization & Advanced Learning (ADR-087)** — Meta-learning enhancements (DecayingBeta, PlateauDetector, ParetoFront, CuriosityBonus), PageRank pattern importance scoring with citation graphs, spectral graph sparsification, reservoir replay buffer with coherence-gated admission, E-prop online learning (RL algorithm #10), and Granger causality for test failure prediction.
+- **456 new tests** across 11 test files covering all Phase 5 milestones with performance benchmarks.
+
+### Fixed
+
+- **HDC pre-filter dropping search results** — The R1 HDC fingerprint pre-filter in PatternStore search was eliminating candidates with low Hamming similarity instead of just reordering them, causing text searches to return empty results when all candidates had fingerprints.
+
+### Changed
+
+- Coherence gate module extracted from 930-line monolith into 4 focused modules under 500-line limit.
+- `RLAlgorithmType` union extended from 9 to 10 members (added `eprop`).
+- Xorshift128 PRNG extracted to shared utility for reuse across RuVector modules.
+- ADR-087 status promoted from Proposed to Active (milestones 1-4 complete).
+
+### Security
+
+- Bumped `path-to-regexp` from 0.1.12 to 0.1.13 (dependabot).
+
+## [3.8.11] - 2026-03-27
+
+### Added
+
+- **YAML deterministic QE pipelines** — Declarative, token-free quality gates with approval steps, auto-approve timeouts, and 4 built-in SQL-only actions (quality-gate, coverage-threshold, pattern-health, routing-accuracy). Full CLI support via `aqe pipeline load/validate/run/list/status/approve/reject`.
+- **Heartbeat CLI and MCP integration** — Manage the token-free heartbeat scheduler with `aqe heartbeat status/run-now/history/log/pause/resume` and 3 MCP tools. Includes history persistence and daily log viewing.
+- **Economic routing model** — Quality-per-dollar tier scoring with budget-aware selection, cost-adjusted rewards for the neural router, and EMA-smoothed quality estimates. Opt-in via `enableEconomicRouting()`. CLI: `aqe routing economics/accuracy/metrics`. MCP: `routing_economics`.
+- **Session operation cache** — O(1) fingerprint-based cache for repeated operations, running before HNSW similarity search. SHA-256 fingerprinting with TTL, capacity eviction, and cross-session persistence via SQLite. Delivers 40-60% token savings on repeated tasks.
+
+### Fixed
+
+- **Path traversal in heartbeat log** — CLI `heartbeat log --date` now validates date format (`YYYY-MM-DD`) to prevent directory traversal attacks, matching the MCP handler's validation.
+- **Approval gate cleanup** — Pending approval gates are now properly resolved when a workflow is cancelled, preventing timer leaks and dangling promises.
+- **Economic config validation** — Routing config weights are clamped to valid ranges and normalized to sum to 1.0, preventing score inflation from invalid inputs.
+- **Sort mutation in tier selection** — `selectTier()` fallback paths no longer mutate the original scores array, ensuring consistent `economicScore` descending order in returned results.
+
 ## [3.8.10] - 2026-03-26
 
 ### Fixed