@overengineered-solutions/ai-runtime

Tenant-scoped AI runtime primitives — spend-consent gate, kill switch, tool-loop detector, provider type contracts.

Status — 0.1.0: types + pure gate logic. Concrete provider adapters (Anthropic, OpenAI, Claude Code, Codex) land in 0.2.0.

Install

pnpm add @overengineered-solutions/ai-runtime

Zero runtime dependencies. Bring your own provider SDKs.

Usage

import {
  isApiRuntimeAllowed,
  shouldShortCircuitToolLoop,
  type EventSink,
} from '@overengineered-solutions/ai-runtime';

// 1) Gate every AI call behind the runtime decision.
const gate = isApiRuntimeAllowed({
  tenant: { tenantId },
  killSwitchOn: row.kill_switch_on,
  pausedUntil: row.paused_until,
  quietHours: row.quiet_hours,
  nowIso: new Date().toISOString(),
});

if (!gate.allowed) {
  throw new Error(`AI runtime gated: ${gate.reason}`);
}

// 2) Detect tool-use loops (same (toolName, inputHash) N consecutive times).
const { shortCircuit, nextState } = shouldShortCircuitToolLoop({
  state: priorState,
  latest: { toolName, inputHash },
  tenantId,
  threshold: 3,
  sink,
});

if (shortCircuit) {
  // surface to operator + halt the conversation loop
}

Why pure functions

isApiRuntimeAllowed and detectToolLoop are pure functions over their inputs. Storage / I/O is the consumer's responsibility — fetch the kill switch, paused_until, quiet_hours from your tenant store, then call the gate. This keeps the package driver-agnostic (no Supabase, no specific ORM) and trivially testable.

License

MIT