The token research we are doing

The free tokens are not a pause in the work. They are how the research gets real repositories. We are studying how a coding session spends tokens, and how to get the same edit with fewer of them. You should not have to count. The product should already be doing it.

The questions

Which steps can move to a smaller model without hurting the edit. How small the context can be before the agent misses the bug. When a second try on the everyday model is cheaper and better than jumping to the careful one. Which parts of a session repeat enough for caching to matter. How often a task comes back as a full rewrite when a diff would do.

What we will not pretend

There is no finished percentage to advertise. The method is the thing we can describe: plan with care, write with the everyday model, push search and summaries down, read ranges, keep caching on, prefer diffs. The notes are how Toknary uses tokens, the noisy work, the plan, the reading, the cache, and the diff.

Your review is the other half of the data. What to include is in what to put in a review.