Context window planner

Output and reasoning tokens come out of the same budget as the prompt on current APIs, so a 200,000-token model given a 198,000-token prompt leaves about 2,000 tokens for the answer. The reserve and the safety margin compete with every section you list, not on top of them. Budgets are whole tokens; on overflow the cut amounts below are exact, not rounded advice. The planner does not tokenize your text, price requests, or know each provider's hidden framing overhead; the safety margin is where you absorb that.