Relay

Purpose

Relay injects failures around a synthetic system-of-record adapter, reads authoritative state back after attempted mutations, reconciles intended versus observed effects, and decides whether to complete, retry safely, replan, route to human review, or stop for manual investigation.

What Relay Provides

Relay provides deterministic failure injection, authoritative read-back, reconciliation, idempotency, version checks, recovery decisions, human review, replay, and regression evals around agent actions.

Execution Boundary

The planner proposes intent. Relay owns deterministic execution safety. The enterprise system owns business truth. Authorization, version checks, idempotency, reconciliation, retry safety, and recovery decisions stay deterministic whether the planner is fixture-driven or backed by xAI/Grok.

Core Reliability Terms

Failure injection controls where the adapter fails. Reconciliation compares intended state against real system state after an attempted action. Idempotency prevents duplicate effects when the same logical request is delivered again. Version checks stop stale plans from overwriting newer source-of-record state.

Architecture

Natural-language request
  -> PlannerProvider (deterministic | optional Grok)
  -> typed ActionPlan
  -> PlanValidator + RiskPolicy
  -> ExecutionCoordinator
  -> FaultInjectingAdapter
  -> Synthetic ClaimsEnterpriseAdapter
  -> PostgreSQL source of record
  -> post-action read-back
  -> ReconciliationEngine
  -> RecoveryEngine
  -> persisted trace / review / eval evidence

What Is Simulated

The included claims workflow is synthetic and provider-agnostic. This repository does not integrate with Guidewire and does not claim to model any third-party company's internal systems.

Limitations

Sequential execution, prototype actors, deterministic recovery rules, fixture-driven local planning, local in-process fault injection, and no real payment or settlement side effects.