A sequence diagram of the session. Each vertical bar is one context: a conversation history the model reads. A Skill or Agent call that opens a new context is a message to a new bar, and the dashed message back carries its result. A new context starts with its prompt only and gets a color of its own; one that starts with a copy of its parent's history says so and keeps the parent's color. Rows are in time order, not to scale. To save width, a column holds one context at a time: when a context ends, a later one can reuse its column, so five runs of the same skill one after another sit in one column as five bars. Such a column's header is an outline with one chip per context, in time order and in the color of its bar; a chip without a name repeats the name before it. Next to the top of a context's bar are the tokens it spent over all its model calls, each of which sends the whole history again: new input, input read from the prompt cache (far cheaper), and output. Hover for the exact numbers. Click a message to see the call's arguments and result.