Use when a task risks running out of context window, when a long-running agent session or multi-step workflow must stay coherent across many turns, when token cost or latency climbs with conversation length, or when the user asks about compaction, summarizing conversation history, truncating tool output, KV-cache or prefix caching, cache-friendly prompt ordering, context budgets, or splitting work across sub-agents. Covers compaction, observation masking, cache-aware layout, and context partitioning, both for an agent managing its own session and for code that builds agent systems. Do not use for RAG retrieval design or vector database selection.