mirror of
https://github.com/proffesor-for-testing/agentic-qe.git
synced 2026-09-19 08:45:47 +08:00
900c0a6390
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1.7 KiB
1.7 KiB
v3.6.19 Release Notes
Release Date: 2026-02-22
Highlights
Fixes 6 learning pipeline issues found via brutal honesty audit: event-driven trajectory lifecycle, object return type detection across all test generators, real learning metrics in fleet_status, and 25 new tests covering every fix.
Fixed
- Event-driven trajectory lifecycle — Replaced fragile polling loop with CrossDomainEventRouter event subscriptions, eliminating race conditions and memory leaks
- Object return type misclassified as primitive — All 5 test generators now correctly detect
{ valid: boolean }as object, not boolean (7 code locations) - memory_store persistence check — Removed unnecessary duck-typing; HybridMemoryBackend always persists to SQLite
- memory_usage reported zeros — Now returns real vector and namespace counts from the backend
- fleet_status missing learning metrics — Added
learningfield with real SQLite counts (patterns, experiences, trajectories, vectors, dream cycles) - Auto-team-wiring crash — Guarded against undefined DomainTeamManager during agent spawn in test environments
Added
- queryCount() on UnifiedMemoryManager with SQL-injection-safe table allowlist
- Trajectory event subscription wired into fleet init/dispose lifecycle
- Orphaned trajectory cleanup at startup for stale trajectories (>1 hour)
- ReasoningBank learning system guide — New docs covering the 4-stage pipeline, pattern lifecycle, and embedding architecture
- 25 new unit tests covering all 6 fixes
Changed
- Embedding dimension consolidation — ReasoningBank/PatternStore standardized on 384-dim (all-MiniLM-L6-v2); coverage-analysis 768-dim indices unchanged