# Token awareness

Cost is dominated by everything RESENT around answers: history, scaffolding, tool output — a mechanism problem, not a "cheaper model" problem.

## Where tokens go
Every turn resends history; every tool result is resent forever after; scaffolding is re-paid each round on runtimes without stable prefix caches; scheduled runs replaying their own history multiply all of it.

## Kaptain's mechanisms
- Context prepared up front: KodeGraph hands the model relevant code (measured: /kaptain/docs/kodegraph-benchmark/ — fewer tokens at equal recall).
- Delegation: mechanical output tokens land on local hardware; batch results go to disk, not context.
- Compaction: old history folds into an inspectable summary when a conversation outgrows its window — visibly.
- Fresh-by-default scheduled runs: no replay tax.
- Lean scaffolding for small windows (re-prefilling runtimes pay for every injected line, every round).

## Small windows are real
4B + 8k + tools = one failed round from full; Kaptain wraps up honestly instead of hallucinating a finish. Give small models bigger windows where hardware allows (the gear shows real default/max).

## See the bill
BlackBox > Model usage: exact provider-reported tokens per run/day/month/route.