Error Recovery Skill

Purpose

Establishes a production-grade failure management and recovery framework for LLM agent systems operating under non-deterministic conditions. LLM agents face unique failure modes — stochastic output degradation, tool call failures, context window overflows, rate limiting, hallucination cascades, and partial state corruption — that traditional error handling cannot address. This framework provides structured error classification, intelligent retry orchestration, checkpoint-based state persistence, multi-tier fallback chains, dead letter processing for permanently failed tasks, error budget enforcement for reliability governance, and chaos testing methodologies to proactively discover failure modes before production deployment.

Core Principles

1. Classify Before Recovering: Every error must be classified into the agent error taxonomy before any recovery action is taken. Retrying a non-retryable error wastes resources and compounds failures.
2. Idempotency Is Mandatory: Every retried operation must produce the same side effects regardless of how many times it is executed. Non-idempotent retries cause data corruption and duplicate actions.
3. Checkpoint Everything: Agent state must be checkpointed after every successful step. Recovery must resume from the last checkpoint, never restart from scratch.
4. Degrade Gracefully, Never Silently: When recovery fails, the system must degrade to a lower-capability mode with explicit user notification, never silently produce degraded outputs.
5. Error Budgets Drive Decisions: Reliability targets are expressed as error budgets. When the budget is exhausted, the system must freeze deployments and prioritize stability over features.

Agent Protocol