Deploy on ankka

The service deploys like any ankka service: an image, a descriptor, and a secret.

Build the image

sbt Docker/publishLocal

This builds ankka-experiment-proposals:<version> and :latest into the local Docker daemon. For a local kind installation, load it into the cluster:

kind load docker-image ankka-experiment-proposals:latest --name ankka

For a remote cluster, set DOCKER_REPOSITORY and run sbt Docker/publish.

Create the secret

The descriptor reads the API token and the model key from a Kubernetes Secret named edp-config, which you create in the project's namespace:

kubectl -n <project namespace> create secret generic edp-config \
  --from-literal=api-token=<token> \
  --from-literal=anthropic-api-key=<key>

Apply the descriptor

service.json

{
  "name": "ankka-experiment-proposals",
  "service": {
    "image": "ankka-experiment-proposals:latest",
    "runtime": "0.12.0",
    "minInstances": 1,
    "env": [
      { "name": "EDP_TIER", "value": "live" },
      { "name": "EDP_MODEL", "value": "claude-opus-5-5" },
      { "name": "EDP_EMBEDDING_MODEL", "value": "BAAI/bge-small-en-v1.5" },
      { "name": "EDP_REVIEWERS_FILE", "value": "/ankka/config/reviewers.json" },
      { "name": "EDP_CONTACT_EMAIL", "value": "change-me@example.org" },
      { "name": "EDP_API_TOKEN", "secretKeyRef": { "name": "edp-config", "key": "api-token" } },
      { "name": "ANTHROPIC_API_KEY", "secretKeyRef": { "name": "edp-config", "key": "anthropic-api-key" } }
    ]
  }
}

ankka services apply -f service.json
ankka services expose ankka-experiment-proposals

runtime must match the ankka version in build.sbt; the two move together on an upgrade. The platform provisions the service's database. Set EDP_CONTACT_EMAIL to a real address before searching the external indexes, which ask for one.

Access

/api/v1 and /review require the token; /healthz is open and read-only. Reviewer identity is a name from the configured list, recorded on each decision; the token says who may call, not who they are. Nothing in the application needs more than one instance.

After deploying

Run the walkthrough against the exposed hostname with --base https://<hostname>. In the live tier the first run calls the model at every stage, so it takes minutes rather than seconds.