Enough for the need.
Not everything. The information required to answer, verify, analyze or execute without another expensive investigation.
An answer is assembled to a size, in a set order: what the question named goes in first and is never trimmed, then what those parts reach. When something had to be left out to fit, the answer says how many more parts exist and what kind they were. So a partial answer is never mistaken for the whole picture.
Ask about a change in the workspace and read the parts the answer was assembled from, grouped by area of the product.
Read the line at the end that says what did not fit before you treat the answer as complete.
From an agent, call the context tool over MCP with the token budget you are willing to spend.
Ask again with a larger budget, or name one of the omitted parts directly, when the gap matters.
The size is chosen by whoever calls. In the workspace you cannot type a number — the budget is set for you; only an MCP or API caller picks it. Ask for too little and the answer comes back marked as too small for even the named parts rather than quietly trimmed. And "enough" is a budget rule, not a guarantee of coverage: it promises to tell you what it left out, not that nothing relevant was missed.
See it on a real product.
Three models are published in full and open without an account. Ask one of them the question this page is about.