“Forma” has a simple rule: spend is counted by fact. A node cannot know from the inside how much it spent: that knowledge lives in the call’s metadata, not in its context. So the spend line is written not by the node itself, but by whoever called it, right after the return.
What goes into the line
The spend line is written into the card’s history and looks like this:
Node, date: attempt, N tokens (R cache-read), T s, <call id> — what was done.
- N — tokens of the call;
- R — cache-read, almost always fixed overhead: loading the rules and the role;
- T — duration in whole seconds;
- id — the call identifier: it links the card to the tool log.
The key part of the line is written in English in every project, so that counters read the same way across all projects.
Statistics, not a threshold
There are two thresholds, both in counts: attempts per task and the number of cards per cycle. Tokens and time are statistics: they are recorded by fact, and Core reads them as a trend. They are not set in advance, and a limit is never set in dollars.
When there is no number
If the engine did not return a number, the line is still written in full, and the field gets the marker unknown and the reason is given. Estimating by eye is forbidden. A gap is counted separately and is never passed off as zero: once, zero and “not recorded” looked the same, and a stretch of attempts looked free of charge.
Where the savings are real
The largest saving is built into the design: the schema is written in the model’s native language, which is about 39% of tokens for the same meaning. The saving multiplies with every call of every node until the end of the project’s life.
Comments