Chapter 1 — Controlling AI spend

The daily budget circuit breaker

src/config/settings.py carries a daily_ai_cost_limit (default 10.0 USD). Before every model call the client compares today's running spend against it; once exhausted, the bot stops calling the model and skips trading until the next calendar day. Override it per-run with DAILY_AI_COST_LIMIT in .env.

Set it to a number you would shrug off losing in a week. 2.0 is plenty to learn on.

The "one prediction ate my budget" trap

Cost = (input tokens + output tokens) x price. Two knobs drive it:

• •ai_max_tokens — the output cap per call. The default (8000) is enormous for a probability estimate. Drop it to 1500–2000.
• •The prompt. Long prompts cost more per call. Trim what you pass to the model to the fields that change the answer.

Then watch it: today's spend is in logs/daily_ai_usage.pkl and per-decision cost lands in SQLite and the dashboard. If you cannot see the spend, you cannot control it.

A cheaper model without gutting quality

primary_model takes any OpenRouter slug. The fallback chain only fires on errors — it is not an ensemble, so you pay for exactly one model per decision.

daily_ai_cost_limit  = 2.0     # USD/day, hard stop
ai_max_tokens        = 2000    # output cap per decision
min_confidence_to_trade = 0.55 # only act on the stronger reads

Change one knob, watch it for a day, then change the next.