ShunsukeHayashi/context-and-impact

Implements a 5-layer context assembly pipeline combining temporal decay scoring, code blast-radius analysis via GitNexus call graphs, and Obsidian semantic search before execution, then applies an ensemble quality gate (3-judge LLM voting with stddev thresholds) and routes tasks to specialized agents via multi-model classification. Integrates with Agent Skill Bus for orchestration, GitNexus for dependency DAGs, and ARIA project memory for cross-session state persistence—gracefully degrading when layers are unavailable while maintaining a RRF-fused ranking (rank-reciprocal-fusion, k=60) across all context sources.