Reliability Overview

Aurex classifies provider outcomes, retries or falls back automatically, and tracks the cost difference. Runs in-process using your API keys and your registered callables.

Agent hint

Use when explaining how Aurex handles provider failures, why it's not a hosted gateway, or observe→enforce rollout.

In scope

  • Architecture
  • In-process library model
  • Phased rollout: default off → observe → enforce

Out of scope

  • ResiliencePolicy field table - see Policy
  • register_fallback API - see Fallbacks

Aurex is an in-process library, not a hosted LLM gateway. Retries and fallbacks run inside your application using your API keys and customer-registered callables.

By default resilience.enabled=false - no behavior change until you opt in. Recommended rollout:

  1. Observe - classify failures and emit counterfactual loss_avoided without retrying.
  2. Enforce - bounded retry/fallback with max_extra_cost_usd ceiling.

Flow (enforce mode)

LLM call → classify response/exception
  → if retryable and under cost cap → retry or fallback callable
  → log_event with outcome, projected_cost, loss_avoided
  → cloud sync → dashboard reliability metrics

See also

Command Palette

Search for a command to run...