Run it locally

From zero to a live incident on a real kind cluster.

Prerequisites: Docker, kind, kubectl, uv, Node 20+.

1. Stand up the sandbox cluster

   A local kind cluster with two namespaces — backstop-naive and backstop-hardened — each running a checkout deployment and a protected prod-db.

   kind create cluster --name backstop
   for ns in backstop-naive backstop-hardened; do
     kubectl create namespace "$ns"
     kubectl apply -n "$ns" -f backend/k8s/app.yaml
   done
   
2. Start the backend (uv)

   Three services: the run API, the custom guardrail endpoints, and the read-only infra MCP server.

   cd backend
   cp .env.example .env        # fill in gateway + Bedrock config
   uv sync
   uv run uvicorn backstop.api:app --port 8033
   uv run uvicorn backstop.guardrails.server:app --port 8133
   uv run python -m backstop.infra_mcp
   
3. Start the frontend (npm)

   cd frontend && npm install && npm run dev   # http://localhost:3033/run
   
4. Trigger an incident

   Open the console, hit Trigger incident, and watch the two agents diverge — the naive one takes prod-db to zero, Backstop catches the bad output, re-routes, and heals to error_rate = 0.0.

Key env flags

Flag	Effect
BACKSTOP_LIVE=true	Use the live model on the gateway for re-diagnosis on the re-route and recovery.
BACKSTOP_LLM_JUDGE=true	Enable the independent LLM-as-judge in the quality gate.
BACKSTOP_BACKEND=kind	Drive the real kind cluster (the K8sBackend) instead of the deterministic mock.

Run the test suite with cd backend && uv run pytest — 60 passing in ~2s, fully network-isolated.