Lab 1.4 — Hand the Keys

The GPU is warming Qwen3.8. Your job now is to hand the afternoon developer a launch URL and a short runbook — not a wiki, not Ansible, not a tour of every Kustomize file.

The template is already on GitHub: fjcloud/go-app-template.

Estimated time: 10 minutes
Persona: Platform Engineer
Deliverable: You have read AGENTS.md. You can paste the Dev Spaces URL to Act 2.

What the developer actually gets

Path Why it exists

AGENTS.md

The only runbook OpenCode reads. Generate main.go / go.mod / Dockerfile, then git, ephemeral Git, and oc.

opencode.json

Pins OpenCode to in-cluster Qwen3.8 via MaaS. The API key is $OPENAI_API_KEY (injected). The OpenCode binary is still injected by Dev Spaces.

devfile.yaml

Dev Spaces workspace. Installs the ephemeral Git client.

deploy/ pipeline/ gitops/

Golden-path Kustomize. The model applies it; it must not rewrite it (except the Deployment image: at deploy time).

There is no scripts/ directory and no Ansible. Qwen3.8 follows AGENTS.md.

Step 1: Read AGENTS.md

Clone only if you want it on disk. On a shared workshop cluster you can also open the file on GitHub. Either way, read it — it is short.

git clone --depth 1 https://github.com/fjcloud/go-app-template
cd go-app-template
wc -l AGENTS.md
cat AGENTS.md

You should see a short file. Notice:

  • Files YOU must generate — main.go, go.mod, Dockerfile with an exact two-stage template

  • Files already in the repository — do not rewrite deploy/, pipeline/, gitops/, devfile.yaml, or opencode.json

  • Three deploy phases — ephemeral Git → Tekton PipelineRun → Argo CD Application

  • Namespaces — ${APP_NAME}-build and ${APP_NAME}-dev only. Never namespace default for the app, Argo CD, or the Application.

Phase What it does Typical duration

1. Git

Create a personal repo on ephemeral Git, commit, and push

~30s

2. Image

oc apply -k pipeline/base and a Tekton PipelineRun (git-clone + buildah)

3–5 min

3. GitOps

Point image: at the built image, oc apply -k gitops/base, create the Argo CD Application

2–3 min
Prompt to production

Step 2: Confirm OpenCode is wired to MaaS

OpenCode is injected from the AI tool registry you applied in Lab 1.3. The binary comes from the platform. The model pin is opencode.json in the template — that file wins over ~/.config/opencode.

CHE_NS=$(oc get checluster -A -o jsonpath='{.items[0].metadata.namespace}')

oc get configmap opencode-workspace-config -n "$CHE_NS" \
  -o jsonpath='{.data.opencode\.json}' | python3 -m json.tool

You should see baseURL {env:OPENAI_BASE_URL} (mint + configure-platform.sh set that env to https://maas.<apps-domain>/llm-serving/qwen3/v1), provider/model name Qwen3.8, API id qwen3, limit.context: 262144, and limit.output: 32768.

If you are on the shared workshop cluster, the ephemeral Git URL in devfile.yaml already matches. Do not fork and sed unless you are adapting this template to another cluster.

Step 3: Copy the developer launch URL

DEVSPACES_URL=$(oc get checluster -A -o jsonpath='{.items[0].status.cheURL}')
if [[ -z "$DEVSPACES_URL" ]]; then
  DEVSPACES_URL="https://$(oc get route devspaces -n openshift-operators \
    -o jsonpath='{.spec.host}')"
fi
TEMPLATE_URL="https://github.com/fjcloud/go-app-template"

echo "============================================"
echo " Participant workspace URL:"
echo "  $DEVSPACES_URL/#$TEMPLATE_URL?ai-provider=opencodeai/opencode"
echo "============================================"

When a developer opens that URL, Dev Spaces clones the template (including opencode.json), injects the OpenCode binary, installs the ephemeral Git client, and opens VS Code in the browser. They talk to Qwen3.8; AGENTS.md is the runbook.

Act 1 complete

Component Status

GPU node pool (1× A100 80 GB)

✅ Provisioned on ARO HCP

Operators + dashboard

✅ https://rh-ai.apps.<apps-domain>/

Qwen3.8 as Models-as-a-Service

✅ Gateway + sk-oai- key; model finishing load in the background

Go app template

✅ Short AGENTS.md; Dev Spaces URL ready to paste

Switch persona — move on to Act 2: first day (Lab 2.1).