horizon

Context window management in long-running LLM sessions involves monitoring context health and compressing stale reasoning. Horizon tracks context budget states and distills subagent context to prevent silent quality degradation as interactions extend.

Reasoning quality degrades in extended multi-turn interactions because the context window fills with stale data. A context compression protocol reduces bloat while preserving essential decisions to maintain performance.

The best way to hand off context to a fresh LLM session is to generate a clean handoff package. This ensures seamless transitions by carrying over only the essential, compressed context needed for continued performance.

To optimize subagent context for long-running tasks, apply a context compression protocol that keeps only the necessary context for subagents. This prevents bloat and maintains processing performance throughout the session.

Trigger a context compression protocol in a long chat when context health monitoring detects Yellow or Red states. These alerts indicate the context window is filling up and immediate compression is needed to preserve reasoning quality.