mirzahusadzic/cogx: v2.7.4 - Service-Oriented Transformation & Intelligence

This release marks a major architectural milestone: the transition from a monolithic "Utility Soup" to a robust, service-oriented architecture. It includes a complete overhaul of core utilities, the introduction of formal state machines for session management, a deconstructed Grounded Context Pool (PGC), and granular architectural intelligence via a project-aware role classification system.

🚀 New Features

• 🧩 Architectural Granularity — Implemented a JSON-based configuration system for path-aware roles and architectural intelligence.
• 🧹 Patterns Compaction CLI — New patterns compact command to reclaim storage space in .open_cognition.
• 🤖 /onboard-roles Command — Automated architectural role classification for project onboarding.
• 🛡️ Tool Output Safety Limits — Strict token-aware safety limits for tool outputs to prevent context flooding.

🧹 Infrastructure & Cleanup

• ⚙️ Formal Session State Machine — Robust state machine to manage session lifecycles and transitions.
• 📦 PGC Deconstruction — Deconstructed PGCManager into specialized, single-responsibility services.
• 📡 Domain/Network Model Separation — Strict boundary between ZeroMQ IPC messages and domain models.
• 🔑 Unique Manifest Keys — Migrated overlays to use unique keys for precise symbol tracking and lineage.

🐛 Bug Fixes & Stability

• 📉 Storage Bloat Remediation — Fixed major storage bloat in .open_cognition with automatic compaction.
• 🧱 Clean Slate Protocol — Protocol to purge tool logs and reset internal state between tasks for context hygiene.
• 🩹 Manifest & Vector Stability — Fixed manifest staleness and optimized orphaned vector cleanup.
• 📡 Vertex SDK Resilience — Added auto-retry logic for JSON parse errors in Vertex SDK responses.

⚠️ Migration Note

It is highly recommended to run a full rebuild of your overlays to take advantage of the structural and lineage improvements: cognition-cli overlay generate --force

Impact

This release significantly improves the stability and performance of the cognition engine, reduces storage footprint by up to 40%, and provides a cleaner, more modular foundation for future intelligent services.