ToknaryThe largest token cost in a coding session is often the reading, not the answer. Pasting a whole file to change ten lines makes the model reread everything it already saw. Toknary is built to search first and then read a range.
Search, then a range
Find the symbol. Open the lines around it. Leave the rest of the file out of the prompt. A short session that starts from the last fact is cheaper than one long transcript that carries every file from the beginning.
What we are measuring
We want to know how small that window can be before the agent misses the bug. Too small, and it edits the wrong place. Too large, and the session pays to reread code that never changes. The free period is when this gets tried on repositories we do not control. Your review, when it says the agent missed a file or read far too much, is the measurement.
Repeated reading of the same prefix is the other half, in prompt caching and repeated work.