Azure App Registration (per repo)

When to use

Invoke this skill when:

• The user asks for "Azure access", "AI Foundry credentials", "Azure OpenAI from this app", "service principal", or similar in the current workstream.
• A Liliput-managed app you are building needs to call Azure AI Foundry, Azure OpenAI, Storage, or AI Search at runtime from its container.
• You see code that reads AZURE_CLIENT_ID / AZURE_CLIENT_SECRET / AZURE_TENANT_ID and they are missing in the dev-env.

Do NOT invoke this skill when:

• The user only needs Liliput itself (the orchestrator) to talk to Azure — Liliput already runs as a managed identity.
• The current task does not need Azure data-plane access. Don't pre-provision credentials "just in case".

What it does

For the repo currently being worked on (e.g. acme/widgets):

1. Finds (or creates) an Entra ID app registration named liliput-{env}-{owner}-{repo}, double-keyed by Liliput-owned tags so it cannot collide with manually-created apps.
2. Ensures the app has a service principal in your tenant.
3. Mints a fresh client secret if the current one expires within 7 days (default 30-day lifetime). Otherwise reuses the existing secret value (kept in liliput/azure-sp-{owner}-{repo} as the source of truth).
4. Assigns the SP a default role bundle on the resource scope given by LILIPUT_AI_FOUNDRY_SCOPE:
   • Cognitive Services OpenAI User
   • Azure AI Developer (AI Foundry user)
   • Azure AI Inference Deployment Operator (AI Foundry contributor)
   • Storage Blob Data Contributor
   • Search Index Data Contributor
5. Writes a Kubernetes Secret named liliput-azure-sp into the dev-env namespace with these keys: AZURE_TENANT_ID, AZURE_CLIENT_ID, AZURE_CLIENT_SECRET, plus AZURE_AI_FOUNDRY_ENDPOINT / AZURE_OPENAI_ENDPOINT / AZURE_AI_PROJECT_ENDPOINT if configured for this Liliput instance.

The dev-pod consumes the secret via envFrom: secretRef: name: liliput-azure-sp. The Deployment manifest you generate must reference it.

How to call

Loopback-only HTTP endpoint inside the orchestrator pod. From a bash step:

curl -sS -X POST \
  -H "Content-Type: application/json" \
  -H "X-Liliput-Internal: $LILIPUT_INTERNAL_TOKEN" \
  http://127.0.0.1:5002/api/azure/app-registration/ensure \
  -d '{
    "repo": "acme/widgets",
    "namespace": "dev-acme-widgets-main"
  }'

Response JSON: appId, appObjectId, servicePrincipalId, tenantId, secretName, rotated, expiresAt, rolesAssigned. The clientSecret value is never returned over HTTP — it lands in the Kubernetes Secret only.

Wiring the secret into the deployment

spec:
  template:
    spec:
      containers:
        - name: app
          envFrom:
            - secretRef:
                name: liliput-azure-sp

In the app code, DefaultAzureCredential picks up the three env vars automatically:

import { DefaultAzureCredential } from '@azure/identity';
const credential = new DefaultAzureCredential();

Optional parameters

{
  "repo": "acme/widgets",
  "namespace": "dev-acme-widgets-main",
  "roleAliases": ["cognitive-services-openai-user", "ai-foundry-user"],
  "scope": "/subscriptions/<sub>/resourceGroups/<rg>",
  "extraSecretData": { "AZURE_AI_PROJECT_NAME": "my-project" },
  "forceRotate": false
}

Failure modes

• LILIPUT_AI_FOUNDRY_SCOPE unset → tool refuses with clear error. Ask the user to set it (ARM resource ID).
• Conflicting app exists without Liliput tags → tool refuses to mutate. User must rename/delete the conflicting app or change LILIPUT_ENV.
• Graph permission denied → Liliput's MI is missing Application.ReadWrite.OwnedBy. See docs/azure-permissions.md.
• ARM permission denied → Liliput's MI lacks User Access Administrator at the configured scope. Same doc.
• Endpoint returns 503 → LILIPUT_INTERNAL_TOKEN not set. Ask user to set it via the deploy workflow.

Important constraints

• One app reg per repo (not per branch). Branches share the SP.
• Secret value is only known at mint time. If a valid secret exists in the central record, it is reused; the tool does NOT rotate just because a namespace is new.
• Rotation does not auto-restart pods. When rotated: true, suggest kubectl rollout restart deploy/<name> -n <ns> to consumers.
• Never log the response body verbatim — treat the endpoint as credential-issuing even though it does not return the secret today.