Agent Ops Lab — Environment

Docker Compose sandbox that simulates a typical internet-company microservices architecture, with injectable faults and recorded ground-truth answers, used to train and evaluate an SRE/RCA (root-cause-analysis) troubleshooting agent.

Related repos

• spring-petclinic-microservices (fork) — business services (Customers/Vets/Visits/Gateway), modified with chaos toggles
• ops-agent-toolkit-mcp — MCP tool server the agent uses to query this environment (logs, metrics, traces, service health/config)

Architecture

Frontend → API Gateway → Customers / Vets / Visits (Java, Spring Boot)
                              │
                    Consul (discovery + config KV)
                    PostgreSQL, Redis
                    Prometheus + Grafana, Loki, Jaeger

See ROADMAP.md for the full phased plan, component rationale, and architecture diagrams.

Quick start

git clone <this-repo>
cd lab-environment

./scripts/build.sh           # build service images from the petclinic fork
docker-compose up -d
./scripts/init-consul-kv.sh  # seed base config — Consul runs in -dev mode, non-persistent

Then open:

Service	URL
App	http://localhost:180
Grafana	http://localhost:3100
Jaeger	http://localhost:16786
Consul	http://localhost:8600

Trigger a fault

./scripts/inject-chaos.sh redis_connection_timeout

Scenarios and their ground truth are defined in scenarios/scenarios.yaml.

Run a full evaluation

./scripts/run-eval.sh

Runs every scenario in sequence, triggers the agent, and scores its conclusion against ground truth. Results land in eval/results/.

Reset everything

docker-compose down -v

Full plan

See ROADMAP.md.