agentic-qe 3.7.7 → 3.7.8

package/.claude/helpers/statusline-v3.cjs

@@ -308,7 +308,19 @@ function getLearningMetrics(projectDir) {
   // V3 trajectories (new V3 trajectory tracking)
   const trajectories = parseInt(sqlite3Query(dbPath, 'SELECT COUNT(*) FROM qe_trajectories')) || 0;
   // Captured experiences (task execution captures)
-  const capturedExp = parseInt(sqlite3Query(dbPath, 'SELECT COUNT(*) FROM captured_experiences')) || 0;
+  // Use SUM(consolidation_count) for monotonically non-decreasing counter:
+  // - New experience → +1
+  // - Merge A into B → A excluded (consolidated_into set), B's count += A's count → net 0
+  // - Archive → row stays, still counted → net 0
+  // Falls back to COUNT(*) if consolidation columns not yet added
+  let capturedExp = 0;
+  const consolidatedQuery = sqlite3Query(dbPath,
+    "SELECT COALESCE(SUM(consolidation_count), COUNT(*)) FROM captured_experiences WHERE consolidated_into IS NULL OR consolidated_into = 'archived'", '__FAIL__');
+  if (consolidatedQuery !== '__FAIL__') {
+    capturedExp = parseInt(consolidatedQuery) || 0;
+  } else {
+    capturedExp = parseInt(sqlite3Query(dbPath, 'SELECT COUNT(*) FROM captured_experiences')) || 0;
+  }
   // Memory entries with learning data (MCP-stored experiences)
   const memoryLearning = parseInt(sqlite3Query(dbPath, "SELECT COUNT(*) FROM memory_entries WHERE key LIKE 'learning%' OR key LIKE 'phase2/learning%'")) || 0;
   // QE pattern usage (hook-recorded outcomes from aqe hooks post-task/post-edit)

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

@@ -904,7 +904,7 @@
   },
   "metadata": {
     "generatedBy": "Agentic QE Fleet",
-    "fleetVersion": "3.7.7",
+    "fleetVersion": "3.7.8",
     "manifestVersion": "1.3.0",
     "lastUpdated": "2026-02-04T00:00:00.000Z",
     "contributors": [

package/CHANGELOG.md

@@ -5,6 +5,25 @@ 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).
 
+## [3.7.8] - 2026-03-04
+
+### Added
+
+- **Loki-Mode adversarial quality gates (ADR-074)** — 7 new features to catch sycophantic AI outputs, hollow tests, and routing drift. All enabled by default (opt-out via config flags):
+  - **Anti-sycophancy scorer**: Detects rubber-stamp consensus via Jaccard similarity, confidence uniformity, and reasoning overlap across model votes
+  - **Test quality gates**: Catches tautological assertions (`expect(true).toBe(true)`), empty test bodies, and missing source imports in generated tests
+  - **Blind review orchestrator**: Runs N parallel test generators with varied temperatures, deduplicates results via Jaccard similarity
+  - **EMA calibration**: Exponential moving average tracks per-agent accuracy and derives dynamic voting weights, with SQLite state persistence
+  - **Edge-case injection**: Queries historical patterns from the learning store and injects proven edge cases into test generation prompts
+  - **Complexity-driven team composition**: Maps 8-dimension complexity analysis (AST + security + concurrency + API surface) to agent team composition
+  - **Auto-escalation tracker**: Consecutive failures auto-promote agent tier; consecutive successes auto-demote for cost optimization
+- **Smart experience consolidation** — Replace destructive pruning with intelligent consolidation that preserves high-value learning patterns while managing memory growth
+- **Multi-language test generation plan** — Architecture decision records (ADR-075 through ADR-079) for unified test framework type system, Tree-sitter WASM parser, compilation validation loop, backward-compatible API, and language-specific path resolution
+
+### Changed
+
+- **Loki-mode features enabled by default** — All 6 config flags (`enableSycophancyCheck`, `enableTestQualityGate`, `enableEdgeCaseInjection`, `enableEMACalibration`, `enableAutoEscalation`, `enableComplexityComposition`) default to `true` for immediate quality improvement
+
 ## [3.7.7] - 2026-03-02
 
 ### Added

package/README.md

@@ -94,6 +94,7 @@ For client-specific setup examples, see [Platform Setup Guide](docs/platform-set
 - ✅ **OpenCode Support** (v3.7.1): 59 agent configs, 86 skill configs (78 QE + 8 general dev), 5 tool wrappers, SSE/WS/HTTP transport, output compaction, graceful degradation, `aqe init --with-opencode` auto-provisioning
 - ✅ **AWS Kiro Support** (v3.7.2): 87 agent configs, 86 skill configs, 5 event-driven hooks, 2 steering files, MCP config, `aqe init --with-kiro` auto-provisioning
 - ✅ **Multi-Platform Support** (v3.7.4): 8 new platform integrations — GitHub Copilot, Cursor, Cline, Kilo Code, Roo Code, OpenAI Codex CLI, Windsurf, Continue.dev — with JSON/TOML/YAML config generation, behavioral rules, and `aqe platform list/setup/verify` CLI
+- ✅ **Loki-Mode Quality Gates** (v3.7.7): Anti-sycophancy scoring, test quality gates, blind review, EMA calibration, edge-case injection, complexity-driven composition, auto-escalation — enabled by default (opt-out)
 - ✅ **V2 Backward Compatibility**: All V2 agents map to V3 equivalents
 - ✅ **78 QE Skills**: 46 Tier 3 verified + 32 additional QE skills (QCSD swarms, n8n testing, enterprise integration, qe-* domains)
 
@@ -299,6 +300,41 @@ aqe learning stats
 
 ---
 
+### 🛡️ Loki-Mode Quality Gates (v3.7.7)
+
+V3.7.7 adds **7 adversarial quality features** inspired by loki-mode — designed to catch sycophantic AI outputs, hollow tests, and routing drift. All features are **enabled by default** (opt-out via config flags).
+
+| Feature | Config Flag | Description |
+|---------|------------|-------------|
+| **Anti-Sycophancy Scorer** | `enableSycophancyCheck` | Detects rubber-stamp consensus via Jaccard similarity, confidence uniformity, and reasoning overlap |
+| **Test Quality Gates** | `enableTestQualityGate` | Catches tautological assertions (`expect(true).toBe(true)`), empty test bodies, and missing source imports |
+| **Blind Review** | N/A (API option) | Runs N parallel test generators with varied temperatures, deduplicates via Jaccard |
+| **EMA Calibration** | `enableEMACalibration` | Exponential moving average tracks per-agent accuracy, derives dynamic voting weights |
+| **Edge-Case Injection** | `enableEdgeCaseInjection` | Queries historical patterns and injects proven edge cases into test generation prompts |
+| **Complexity Composition** | `enableComplexityComposition` | Maps 8-dimension complexity (AST + security + concurrency) to agent team composition |
+| **Auto-Escalation** | `enableAutoEscalation` | Consecutive failures auto-promote agent tier; consecutive successes auto-demote |
+
+```typescript
+// All features are ON by default. To disable specific features:
+const config: Partial<RoutingConfig> = {
+  enableEMACalibration: false,    // Disable EMA voting weights
+  enableAutoEscalation: false,    // Disable auto tier promotion
+};
+
+const consensusConfig: Partial<ConsensusEngineConfig> = {
+  enableSycophancyCheck: false,   // Disable rubber-stamp detection
+};
+
+const testConfig: Partial<TestGeneratorConfig> = {
+  enableTestQualityGate: false,   // Disable tautology detection
+  enableEdgeCaseInjection: false, // Disable pattern injection
+};
+```
+
+> See [docs/loki-mode-features.md](docs/loki-mode-features.md) for detailed usage examples and configuration reference.
+
+---
+
 ### 🌙 Dream Cycles & Neural Learning
 
 V3 introduces **Dream cycles** for neural consolidation and continuous improvement: