ToknaryThe plan is where a coding agent either understands the task or wanders. Toknary spends its careful pass there. The writing that follows uses the everyday model, at a low effort, because the direction is already chosen.
When the careful pass returns
If the everyday model fails the same task twice, the careful pass comes back for another look. It does not sit on every keystroke. High effort is for the plan and for that second failure, not for formatting a file the session has already understood.
What this saves
Most tokens in a session are the middle: reading a range, applying a diff, fixing a test. Those do not each need the strongest setting. Keeping them on the everyday model is how a long task stays affordable even when, during this period, the tokens are free to you. We still want the habit, because the research is about the work, not about a bill.
The wider split is in how Toknary uses tokens.