Run the Local Service Stack

Bring up the whole service-driven flow — backend (run queue), auth, dispatcher, and artifact service — on a local k3d cluster (the web console runs from source, not in-cluster), and drive runs the way the web console does in the cloud: enqueue a run, and an in-cluster dispatcher schedules it as a per-run driver Job. This is the service-driven alternative to the single CLI Run a Test Case.

For the full walkthrough and the why, see Running the Local Service Stack; for every service variable, Running.

Prerequisites

• Docker (or a Docker-compatible runtime), k3d, kubectl, and make on PATH. (k3d’s first-class runtime is Docker; on Podman it needs rootful + the Docker socket.)
• A harness API key exported — ANTHROPIC_API_KEY (claude), OPENAI_API_KEY (codex), or OPENROUTER_API_KEY (cline/goose/kilo/opencode/pi). The Makefile reads it from the environment into the cluster Secret; nothing is written to a tracked file.

Steps

export OPENROUTER_API_KEY=…                   # the harness you'll run

make -C deployments/local local-up            # cluster + images (backend/auth/dispatcher/driver/artifact/arena) + secrets + overlay + ingest
make -C deployments/local local-forward       # backend → :8787, auth → :8789, artifacts → :8790, arena → :8791 (leave running)
npm run -w apps/web dev                        # the web console, from source → :1430 (its own terminal)

The web console runs from source (not in-cluster), pre-pointed at the forwarded backend/auth via the committed apps/web/.env.development — no VITE_BACKEND_URL to set. Open it at http://127.0.0.1:1430 and the catalog loads on first visit. Then, in the console:

1. Register / log in (or tcab register --username dev --display-name "Dev") so push/review/publish are attributed to you.
2. Enqueue a run — pick a case, a model, and the harness whose key you exported. The console streams its events live.
3. Watch it schedule as a Job: kubectl -n tcab-local get jobs,pods -w shows a driver Job and (Kubernetes runtime) a sandbox pod. The finished run is reviewable, its build served from the artifact service.

Manage

make -C deployments/local local-status        # pods, services, volumes
make -C deployments/local local-rebuild       # after a code change: rebuild + restart
make -C deployments/local local-ingest        # after editing a case: force re-ingest
make -C deployments/local local-down          # delete the cluster and everything in it

Next steps

• Running the Local Service Stack — the full guide, including the topology and troubleshooting.
• Review a Run · Publish a Run — what to do with the run you produced.