mirror of
https://github.com/proffesor-for-testing/agentic-qe.git
synced 2026-09-19 08:45:47 +08:00
9ff41861b1
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
1.7 KiB
1.7 KiB
v3.6.17 Release Notes
Release Date: 2026-02-21
Highlights
The QE Queen Coordinator now runs its full 10-phase protocol end-to-end. Semantic search via HNSW actually works — store a value, query with natural language, get scored results back. Hierarchical topology assigns domain leads and workers. MCP connections survive parallel load.
Added
- HNSW semantic search —
memory_querywithsemantic: trueperforms vector search with cosine similarity scoring and automatic fallback to pattern matching - Vector storage on write —
memory_storeindexes entries in the HNSW vector index so semantic search finds them immediately - Hierarchical agent topology — First agent per domain becomes lead, subsequent agents become workers when using
topology: "hierarchical" - MCP reconnection with buffering — Transport errors trigger exponential backoff reconnection with request buffering and replay
- Real experience capture — Tool executions record patterns via ExperienceCaptureService instead of placeholder feedback
Fixed
- Stdio transport drops under parallel load — Write timeout 30s to 120s with retry and backpressure handling
- Semantic search returned 0 results —
memory_storenow callsstoreVector()after every KV write - Embedding dimension mismatch — Text embeddings now generate 768 dimensions matching the HNSW index
- Transport stop() race condition —
runningflag set before readline close to prevent spurious reconnect
Changed
- Queen Coordinator docs updated with semantic search params and complete tool reference
- Learning feedback reflects actual capture status instead of hardcoded text