proffesor-for-testing__agentic-qe/docs/releases/v3.7.8.md

v3.7.8 Release Notes

Release Date: 2026-03-04

Highlights

Introduces Loki-Mode adversarial quality gates (ADR-074) — 7 features that catch sycophantic AI consensus, hollow test generation, and static routing drift. All enabled by default for immediate quality improvement.

Added

• Anti-sycophancy scorer — Detects rubber-stamp consensus via 4 weighted signals: verdict unanimity, Jaccard reasoning similarity, confidence uniformity, and issue count consistency. Triggers Devil's Advocate review on severe sycophancy.
• Test quality gates — Structural validation catches tautological assertions, empty test bodies, missing source imports, and mirrored assertions in AI-generated tests. Each test gets a quality score (0-100).
• Blind review orchestrator — Runs N parallel test generators with varied temperatures, deduplicates via Jaccard similarity on tokenized assertions, returns merged set with uniqueness statistics.
• EMA calibration — Exponential moving average tracks per-agent success rate and derives dynamic voting weights (floor 0.2, ceiling 2.0). State persists to SQLite across sessions.
• Edge-case injection — Queries historical patterns from learning store, ranks by success rate, and injects top-N proven edge cases into test generation prompts before LLM calls.
• Complexity-driven team composition — Analyzes code across 8 dimensions (AST metrics + security/concurrency/data-flow/API surface) and maps to agent team composition. High security score adds security auditor; high concurrency adds chaos engineer.
• Auto-escalation tracker — 3 consecutive failures auto-promote agent tier (haiku -> sonnet -> opus); 5 consecutive successes auto-demote for cost optimization. Uses existing fallback chain.
• Smart experience consolidation — Replaces destructive pruning with intelligent consolidation that preserves high-value learning patterns.
• Multi-language test generation ADRs — Architecture decisions (ADR-075 through ADR-079) for unified type system, Tree-sitter WASM parser, compilation validation, backward-compatible API, and language-specific path resolution.

Changed

• All 6 loki-mode config flags default to true (opt-out instead of opt-in)

Architecture

• 14 new source files, 7 test files, 178 dedicated tests
• ADR-074 documents all decisions and trade-offs
• See docs/loki-mode-features.md for detailed usage guide