Cost management
20 September 2026
As much as I would like this experiment to be a huge free sandbox, it is not. Using AI models costs money — some more than others, depending on how capable they are.
And with Woodhouse being allowed to act slightly autonomously, I will need to keep one eye on cost management.
I am already focused on choosing the right model for the right task. This configuration focuses on token usage instead.
What configuration can I make that will reduce token usage? How do I avoid sending input tokens that are not needed? And how can I make better use of cached tokens?
Where this stands
There are no spending controls at all. Not a low limit — none. No token cap, no budget, no spend alert, and no rule that stops a task retrying forever. I went looking for the setting to adjust and there was nothing to adjust.
The only thing standing between me and a runaway bill is a per-job runtime timeout: sixty seconds on each of the entrance-light jobs, four minutes on the morning brief. That limits how long a job runs. It does not limit what it spends while running, and it does nothing at all for conversation.
One saving has already happened, as a side effect of something else. Trimming the instruction files from 232 lines to 101 removed guidance for tools this setup does not have. That is a cost change as much as a tidying one, and it is the kind that compounds: those files are re-read at the start of every session, forever. A single wasteful API call costs once. A wasteful instruction file costs every time the assistant wakes up.
What goes in
Three things, in order of how much they are worth.
Context discipline first. Everything loaded at startup is paid for on every session: the instruction files, and MEMORY.md. Keeping those lean is not housekeeping, it is the recurring bill. This is where the memory configuration and this one meet.
Then hard limits. My own note says a task that fails three times should stop, and nothing should run indefinitely. Neither rule exists in the configuration yet.
Then measurement. I cannot manage what I am not looking at, and right now I have no number in front of me at all.
Model choice is the other half of this and gets its own entry — this one is about the meter rather than which engine is running.