What components count as "context" for Claude on a given turn?	The system prompt, message history, tool definitions/results, attached files, and thinking blocks.
What is the primary risk associated with the context window?	If the window fills, the API request fails.
What is the "Just-in-time context" pattern?	Loading only what the agent needs immediately, and letting tools pull in additional information when required.
How does "Server-side compaction" manage long conversations?	The API automatically summarizes old turns into a single block when the input crosses a trigger threshold, requiring a `context_management` key.
What is the main benefit of using "Prompt caching"?	Reusing stable parts of a request (like system prompts) across multiple calls at a fraction of the cost.
What is the purpose of the "memory tool"?	To allow context (like user preferences or running notes) to survive across different sessions.
Which context management pattern is described as a design pattern rather than a first-class API feature?	Just-in-time context.
What three failure modes do the four context management patterns help address?	Cost, window size limits, and statelessness.