Azure Databricks dependency and configuration review

Use ONLY generate-databricks-dependency-report. The operator must explicitly select that one tool and this skill for databricks-dependency-live; no inherited tools, shell, exec, generic Python, SQL execution tools, remediation connectors or role-grant tools.

Scope, cross-tenant identity and evidence boundaries

The code contains fixed administrator constants: SUBSCRIPTION_ID, SRE_TENANT_ID, TARGET_TENANT_ID, MANAGED_IDENTITY_CLIENT_ID, optional FEDERATED_APP_CLIENT_ID, optional KEYVAULT_OAUTH_SECRET_URI, and the APPROVED_WORKSPACES allowlist. Tool parameters cannot change credentials, tenants, workspace hosts, endpoints, SQL/KQL queries, filesystem paths, billing scopes or resource IDs.

• If the fixed configuration still contains <PLACEHOLDER> values, the tool stops before any authentication or network call with status="error" and error.code="configuration_required" (artifact_status="not_written").
• If SRE_TENANT_ID != TARGET_TENANT_ID and auth_mode == "same_tenant_uami", the tool stops before authentication with error.code="cross_tenant_configuration_required". A single-tenant SRE UAMI in Tenant A never automatically obtains Tenant B tokens or Databricks workspace permissions.
• Keep Azure Resource Manager (ARM) control plane (management.azure.com) strictly separate from the Azure Databricks Workspace & Unity Catalog REST plane (https://adb-<id>.<shard>.azuredatabricks.net). Azure Lighthouse delegation projects ARM RBAC only; it does not grant Databricks Workspace or Unity Catalog access.
• Supported cross-tenant modes are:
  1. cross_tenant_multitenant_fic: Multi-tenant Entra App Registration in the SRE tenant trusting the SRE UAMI via Federated Identity Credential (api://AzureADTokenExchange) and provisioned as a Service Principal in the target tenant for both ARM and Databricks (2ff814a6-3304-4ab8-85cb-cd0e6f879c1d/.default).
  2. cross_tenant_lighthouse_plus_databricks_oidc: Azure Lighthouse for ARM reads plus Databricks Service Principal Federation Policy (/oidc/v1/token RFC 8693 grant_type=urn:ietf:params:oauth:grant-type:token-exchange with client_id, subject_token, subject_token_type=urn:ietf:params:oauth:token-type:jwt, scope=all-apis) for Workspace/Unity Catalog REST reads.
  3. cross_tenant_lighthouse_plus_keyvault_oauth_m2m: Azure Lighthouse for ARM reads plus in-memory OAuth M2M secret resolution from an approved Key Vault in the SRE tenant using the SRE UAMI (https://vault.azure.net/.default).

Use operation="list_workspaces" with other defaults to list the fixed operator-configured roster without authentication or network access. This roster is configuration metadata, not proof that a workspace exists, is reachable or authorizes reads.

Deep dependency mapping contract (OBSERVED vs INFERRED vs UNKNOWN)

Every dependency node, edge and blind spot carries an explicit classification:

• OBSERVED: Exact native ID or key join verified across collected API responses (job_task.depends_on, job_task.existing_cluster_id, job_task.job_cluster_key, job_task.pipeline_id, job_task.warehouse_id, job_task.run_job_id, cluster.policy_id, cluster.instance_pool_id, uc_external_location.credential_name, uc_storage_credential.access_connector_id, secret_scope.keyvault_metadata.resource_id, dlt_pipeline.catalog/target, and ARM customVirtualNetworkId, accessConnector.id, managedResourceGroupId, outboundNetworkDependenciesEndpoints).
• INFERRED: Explicit URI or hostname reference in configuration where the target Azure resource detail was not collected in ARM (for example, an abfss://container@account.dfs.core.windows.net external location or CMK keyVaultUri whose target storage account or Key Vault is outside the collected ARM inventory). Never present an INFERRED reference as a verified live connection.
• UNKNOWN: Uncollected or out-of-scope dependency layers that require running compute or code inspection:
  • Runtime table and column lineage (system.access.table_lineage, system.access.column_lineage) and historical DBU billing (system.billing.usage), which require executing SQL on a running SQL Warehouse.
  • Dynamic paths, JDBC strings or table names embedded inside notebook, Python or JAR code.
  • Effective data-plane packet flow, DNS resolution and storage/Key Vault firewall ACL evaluation.

Never fabricate table lineage from job or notebook names, and never start a cluster or SQL Warehouse to fill gaps.

Report and four-file delivery

Make one complete call with operation="report" and other defaults except an approved workspace_name or workspace_names. No local filter or paging combination is valid for a report. Immediately render all four provided files: HTML (report_path), full normalized JSON (snapshot_path), failed findings CSV (findings_csv_path) and dependencies CSV (dependencies_csv_path).

• For sharing_mode="native_ui", render all four unchanged *_url values (report_url, snapshot_url, findings_csv_url, dependencies_csv_url) as native Markdown links/file widgets in the same response; the tool derives them from the verified SRE runtime thread directory. Never invent or reuse a host/thread/sandbox prefix. http_verified=false does not suppress these provided links; artifact_status="written" is not an HTTP download receipt.
• For sharing_mode="local_files_only", expose the unchanged actual *_path values and state that cwd is outside the supported native runtime roots.
• On status="error", report the sanitized error.code, error.source_id and error.http_status and stop. Do not claim completion, invent files, narrow scope, use ZIP/storage fallbacks or substitute a simulated report.

Grounded questions

Use operation="query" with view in summary, workspace, compute, jobs, pipelines, warehouses, unity_catalog, resources, findings, finops, coverage, actions, or dependencies.

• Use classification (OBSERVED, INFERRED, UNKNOWN) with view="dependencies" to filter the dependency graph by provenance.
• Use workload_name for exact local filtering by cluster, job, pipeline, warehouse, Unity Catalog object or resource name.
• Use area, level, severity and status only where documented.
• Honor exact total_matched, returned, has_more, next_offset, transport_capped and fresh-collection disclosure under the 3,500-byte default-serialized response budget.
• Scores are local assessment indicators, never an official Microsoft WAF or Databricks grade. Always report evaluated, not_evaluated (No evaluado) and not_applicable (No aplica) counts alongside any score.

Never do

No Azure RBAC or Unity Catalog grants, no multi-tenant consent or federation policy creation, no cluster or SQL Warehouse start/stop, no job or pipeline runs, no /api/2.0/sql/statements execution, no /api/2.0/secrets/get calls, no export of spark_env_vars, spark_conf, init script bodies, notebook code or job parameter values, and no fabricated savings or lineage.