Auto-Integration Loops for Capability Self-Propagation

Decision

Implemented 4 closed loops to ensure new capabilities (MCP servers, skills, tools) automatically propagate across the entire CIOS system without manual intervention: (1) Session-close creates capability_change Brain handoffs to all 3 orchestrators; (2) Sentinel runs capability inventory every 30min comparing against Brain baseline; (3) Homeostasis reconciles registration parity and Brain freshness every 4h; (4) All orchestrators query Brain for capability-inventory at session start for dynamic discovery.

Rationale

Gap analysis revealed agents had ACCESS to new tools (via volume mounts) but no AWARENESS of them. Orchestrators used hardcoded capability registries that couldn’t discover new tools. No mechanism existed to bridge Build→Notify→Discover→Route. Solution uses existing Brain handoff mechanism — zero new infrastructure, purely instruction-level changes to Sentinel, Homeostasis, all 3 orchestrators, and session-close skill.

Alternatives Rejected

Considered: (A) New dedicated ‘Registry Agent’ — rejected, adds complexity, existing agents already run periodically. (B) Polling-based approach where every agent scans filesystem on every run — rejected, wasteful, better to have Sentinel centralize discovery and broadcast via handoffs. (C) Webhook/event-driven Docker events — rejected, over-engineered for current scale, cron-based heartbeats sufficient.

Outcome

Pending