Cost and performance guardrails

Keep development inference spend, tokens, and latency visible.

arka guardrails status
arka guardrails reset

Set optional local limits in .env:

ARKA_COST_LIMIT_USD=1.00
ARKA_TOKEN_LIMIT=100000
ARKA_LATENCY_LIMIT_MS=15000

Hosted requests are blocked before dispatch when the token or estimated cost budget is exceeded. Local-only models are never blocked by cost guardrails. Arka stores only aggregate counters in its cache, not prompts.

Grounded mode is enabled by default (ARKA_GROUNDED_MODE=1). It blocks unrequested mock data and instructs model-facing requests to use symbolic tools or configured APIs for current information. Unless you explicitly ask for a mock, prototype, placeholder, or example design, Arka biases toward real data, real integrations, and production-shaped error handling. Disable only when prototyping:

ARKA_GROUNDED_MODE=0 arka ask "..."
arka grounding

Production mode is also enabled by default (ARKA_PRODUCTION_MODE=1). Coding requests are instructed to include access control, secret-safe data handling, input and output validation, prompt-injection defenses, audit logging, tests, error handling, and observability. Disable it only for explicitly disposable prototypes.

Data minimization defaults to schema-only mode (ARKA_DATA_MODE=schema). JSON payloads sent to models are summarized as field names/types or array counts; literal trade-secret values remain available to local tools but are not exposed to the model. Use ARKA_DATA_MODE=full only when the data is safe to share.