Context window

The context window is the total span of tokens the model can see at once: system prompt, tool definitions, CLAUDE.md, skill descriptions, conversation history, file contents and tool results. Each Claude Code session begins with a FRESH context window — nothing carries over automatically, which is why CLAUDE.md and auto memory exist as mechanisms to carry knowledge across sessions. Everything competing for that space is why progressive-disclosure matters: skill descriptions are always loaded, but bodies only load on invocation, and once loaded a skill body stays in context across turns as a recurring cost. Subagents are the other lever: each subagent runs in its OWN context window with its own system prompt and tool access, does the noisy work (search results, logs, file contents), and returns only a summary to the main conversation. When context fills up, auto-compaction summarizes the conversation to free space. For example, Compact Guard helps manage state within Claude Code's real context-compaction limits. When the context window fills, the system may automatically summarize the conversation to make space. This is also known as the context limit or context length, and running out of context is a common operational challenge.