Context compaction

Read Context compaction is the set of protocol and runtime patterns used to keep an AI agent's working context window bounded across multi-turn tool invocations—preventing attention degradation and redundant token spend. On the network side, agentic protocols support compaction by serving Markdown twins (Accept: text/markdown instead of raw DOM HTML), exposing compact: true projections on verbose discovery/probe tools, and isolating noisy exploratory searches inside disposable subagents that return only structured summaries to the parent trajectory. Used by: mcp, llms-txt, cloudflare-markdown-for-agents.