Tools
What the model can use. A change applies from your next message, in the same conversation.
Features
System prompt
The instructions the model gets before the conversation. A saved change applies from your next message, in the same conversation; sub-agents get it too, with their own instructions added. With Claude, a change makes the next request re-read the conversation without the cache once.
Context
How full the model's context window is, and compaction: the conversation so far is summarized by the same model and the summary takes its place. Files in the VM are not affected.
Token usage this conversation
Input counts uncached input tokens only; cached tokens are under Cache read. Output includes thinking. Claude doesn't report thinking tokens separately (—); OpenAI-style APIs that do show them. Sub-agents are the Delegate tool's; Compaction is the requests that wrote summaries.
/agent/skills. Type @id in a prompt to include a skill.