Token Budget

Token budget

ARGUS is frugal by design — every task has hard ceilings; exceeding them stops the run.

Configured in argus.config.json → budget:

Limit	Typical default	On exceed
`maxUsdPerTask`	`$0.50`	`BudgetExceededError` → task ends
`maxTokensPerTask`	`200000`	same
`maxSteps`	`24`	same
`maxToolCalls`	`40`	same

No silent overspend. No "I'll just finish this one more reflection."

Every LLM call records usage + model pricing. Check anytime:

argus status

Shows accrued tokens, USD estimate, steps, tool calls for the current or last task.

Lever	What it does
Model tiering	Triage on cheap/local; escalate to core/heavy only when needed
cache_control	Anthropic ephemeral cache on stable system + tool defs
Compaction	Lessons distilled; context stays bounded
Scoped tasks	"Fix nginx 502" not "analyze every marketplace since 2019"

Agents that never stop billing aren't smart; they're rude.

ARGUS stops when the scoped task is done or when the budget says stop — whichever comes first.