Example: Credential-Gated Deploy to Coolify + Container Registry

The Runner runs on the same Ubuntu 24.04 host as Coolify (matching the setup in docs/self-hosted-runner.md). This means the Runner has inherent network access to Coolify's local API. The credential architecture ensures Agents cannot abuse this access.

# ~/.pks-cli/runner-credentials.yaml   (never mounted into containers)

credentials:
  - service: registry
    endpoint: ghcr.io
    username: pksorensen
    token: "${GHCR_TOKEN}"          # ghp_ PAT with packages:write
    scopes: ["push", "pull"]
    allowed_labels: ["build", "deploy"]

  - service: coolify
    endpoint: https://coolify.internal/api/v1
    token: "${COOLIFY_API_TOKEN}"   # Coolify API key
    scopes: ["deploy", "read"]
    allowed_labels: ["deploy"]
    # production deploys require human approval at the DMZ level:
    require_approval_for_paths: ["/api/v1/deploy", "/api/v1/restart"]

egress_policy:
  # Allow: package registry, GitHub API
  - destination: "ghcr.io"
    allow: true
  - destination: "api.github.com"
    allow: true
  # Require approval: Coolify deploy endpoints
  - destination: "coolify.internal/api/v1/deploy"
    method: POST
    require_human_approval: true
  # Block everything else by default
  - destination: "*"
    allow: false

{
  "name": "Build and Deploy",
  "description": "Build a Docker image, push to ghcr.io, deploy via Coolify.",
  "columns": [
    {
      "id": "build",
      "name": "BUILD",
      "order": 0,
      "trigger": {
        "type": "dispatch_worker",
        "promptTemplate": "Build and push the Docker image for this deployment.\n\n{{task.description}}\n\nSteps:\n1. Read the task description to find: repository URL, image name, tag\n2. Clone the repository\n3. Get registry credentials: curl --unix-socket /run/alp/cred.sock 'http://localhost/token?service=registry&scopes=push'\n4. Log in to the registry using the returned token\n5. Build the Docker image: docker build -t <image>:<tag> .\n6. Push the image: docker push <image>:<tag>\n7. Write the image digest to workspace/image-digest.txt\n8. Call complete_station with conclusion: success",
        "appendSystemPrompt": "You are a build engineer. Use the ALP credential server at /run/alp/cred.sock to get registry credentials — never use hardcoded tokens. Build the image exactly as specified. Write the pushed image digest to workspace/image-digest.txt before completing.",
        "labels": ["build", "docker"],
        "idleTimeoutMinutes": 5,
        "maxTimeoutMinutes": 30
      }
    },
    {
      "id": "deploy",
      "name": "DEPLOY",
      "order": 1,
      "trigger": {
        "type": "dispatch_worker",
        "promptTemplate": "Deploy the built image via Coolify.\n\n{{task.description}}\n\nSteps:\n1. Read workspace/image-digest.txt to get the image digest\n2. Read the task description to find: Coolify application ID, environment\n3. Get a deploy token: curl --unix-socket /run/alp/cred.sock 'http://localhost/token?service=coolify&scopes=deploy'\n4. Trigger the deployment via the ALP egress proxy:\n   curl -X POST https://coolify.internal/api/v1/deploy \\\n     -H 'Authorization: Bearer <jit-token>' \\\n     -H 'Content-Type: application/json' \\\n     -d '{\"applicationId\": \"<app-id>\", \"imageDigest\": \"<digest>\"}'\n   NOTE: The proxy will hold this request for human approval before forwarding.\n5. Wait for the 200 response (it will come after a human approves the deploy)\n6. Log the deployment URL from the response\n7. Call complete_station with conclusion: success and summary including the deploy URL",
        "appendSystemPrompt": "You are a deployment engineer. Use the ALP credential server for all credentials. Your deploy call will be held for human approval — this is expected, just wait for the response. Never bypass the proxy. Record the deployment URL in your summary.",
        "labels": ["deploy"],
        "idleTimeoutMinutes": 60,
        "maxTimeoutMinutes": 90
      }
    }
  ],
  "transitionRules": [
    { "fromColumnId": "build", "toColumnId": "deploy", "condition": "success" }
  ]
}

Title: Deploy my-app v2.3.1 to production

Description:
  Repository:    https://github.com/pksorensen/my-app
  Branch:        main
  Git SHA:       a1b2c3d
  Image name:    ghcr.io/pksorensen/my-app
  Tag:           2.3.1
  Coolify App:   app-abc123
  Environment:   production
  Notes:         Includes the new auth refactor — review the deploy carefully.

The ALP credential model is directly inspired by how GitHub Actions OIDC works with Azure: the CI job never holds a static cloud credential. Its identity (workload identity) is used to request a short-lived, scoped access token at runtime. If the runner environment is compromised, the blast radius is bounded.

This example represents the target architecture. The current state:

Today, the pattern is approximated by injecting credentials as env vars (e.g. ASSEMBLY_LINE_REPO_TOKEN). The credential server is the next evolution, narrowing the blast radius significantly.