Cap your spending¶

By default Reyn runs without any spending limit — it keeps calling the LLM until your task is done. If you want a hard ceiling on tokens or dollars, add a cost: block to reyn.yaml. Caps are checked before each LLM call, so a refused call costs you nothing.

Set a daily dollar cap¶

The most common goal — "never spend more than a few dollars a day":

# reyn.yaml
cost:
  daily_cost_usd:
    hard_limit: 5.00     # refuse the next call once today's spend hits $5
    warn_ratio: 0.8      # warn once at 80% ($4.00)

Daily and monthly caps are persistent — they survive restarts and reset automatically at local midnight (daily) or the 1st of the month (monthly). They are stored in .reyn/state/budget_ledger.jsonl.

Cap a single agent or session¶

Per-agent caps are in-memory — they reset when you restart reyn chat or run /budget reset. Use them to bound one conversation rather than your whole day:

cost:
  per_agent_tokens:
    hard_limit: 50000
  per_agent_cost_usd:
    hard_limit: 2.00

Check where you stand¶

While reyn chat is running:

Command	Shows
`/cost`	One-line spend for the attached agent this session
`/budget`	Full breakdown — today, this month, per-agent, rate limits
`/budget reset`	Clear the in-memory per-agent counters (daily/monthly are untouched)

What happens when you hit a cap¶

At the warn threshold (hard_limit × warn_ratio): you get a one-time [budget warn] message and the call proceeds.
At the hard limit: the next call is refused before it runs. You'll see [budget exceeded] with the triggered dimension and your options — raise the limit in reyn.yaml, run /budget reset, or restart.

Notes¶

Without a cost: block, runs are unlimited — the whole framework is opt-in.
USD figures are estimated from LiteLLM's pricing data. If a model has no price entry the dollar counter stays at $0.00 but token counts are always accurate — so token caps work even when pricing is unknown.
A rate limit (rate_limit_per_minute) refuses calls until the 60-second window clears; Reyn does not auto-sleep, so the caller retries.