What problem does it solve? Choosing which model or backend should handle an inference request is error-prone: teams either overpay for premium models on trivial tasks, silently blow past cost budgets, or hang indefinitely on unresponsive endpoints. This Skill operationalizes a deterministic tier-based routing policy so requests always try the cheapest viable backend first and escalate only with explicit, auditable reasons. ## Core Features & Use Cases - Seven-Tier Routing Hierarchy: Routes requests from in-context and local Ollama/gbrain tiers (free) up through HuggingFace free, proprietary free, OpenRouter paid, and a Grok last-resort tier, matching the production frugality_router.py implementation. - Hard Timeout Enforcement: Mandates 10-second killable background-job timeouts per tier and explicitly forbids the timeout N && cmd pattern that breaks SIGTERM delivery. - Fail-Closed Cost Guard: Integrates CostGuard with a $25/day budget, 80% alert threshold, and raise-on-deny semantics so cost overruns surface as exceptions rather than silent reroutes. - Escalation Audit Trail: Requires an escalation_reason for every tier >= 3 elevation and enforces privacy_critical and offline-mode policies. - Use Case: An orchestrator receives a 3000-token reasoning request; it probes local tiers first, and on timeout escalates to a paid tier only after CostGuard.can_spend() approves, logging the escalation reason for compliance. ## Quick Start Ask the agent to route a reasoning request through the tier hierarchy with cost guard checks and 10-second timeouts, escalating only with a recorded escalation reason.
