Change the lines, do not reprint the file

Output tokens are the lines the model writes back. Reprinting a whole file to change a condition spends those tokens on text you already had. Toknary prefers a diff: the lines that move, and nothing else.

The same edit, less paper

A patch is easier to review in the terminal, and it is smaller to generate. A full rewrite also risks drifting the parts of the file the task never mentioned. The preference is deliberate. Concise answers sit beside it. The session should say what changed, not narrate the file.

Where this meets the research

We are watching how often a task still comes back as a rewrite, and whether that happens because the context was too wide or the instruction was too vague. If your review says the agent rewrote a file it should have patched, that is a direct note on this work. The rest of the spending rules are in how Toknary uses tokens.