Stop a run at a spending cap #14

Priority

extreme

Components

• app/nullius (backend)
• app/web (routes)
• templates and static files
• docs and the site
• catalogue (benchmarks, agents, capabilities)
• the standard (ports, kits, relay, runs)
• CI, workflows and skills

Today

The relay records every call a run makes (app/nullius/standard/relay.py), but nothing stops a run at a budget: a run on a paid key spends until it ends. To run within a fixed budget I put a small proxy between the relay and the provider. Before each call it reserves an upper bound of its cost (prompt characters / 3 as input tokens, max_tokens as output, at list prices). It refuses the call with HTTP 402 when the run's spend plus the calls in flight would pass the cap, and it keeps a ledger of what was spent.

Wanted

A spending cap per run, for anyone running benchmarks on a prepaid or shared key, as an option such as --max-usd 5 or a setting. Once the run's spend, counting the calls in flight, would pass the cap, no further call goes out, and the job says so. A cap per task would be a useful addition, to stop one runaway unit.

Acceptance

• A run can be given a spending cap, and no call goes out once the spend plus the calls in flight would pass it.
• A capped run's result says the cap stopped it, what was spent, and which units did not run.

Dependencies and related

none