import-infrastructure-as-code（将基础设施导入为代码）

Import Infrastructure as Code (Azure -> Terraform with AVM)

Convert existing Azure infrastructure into maintainable Terraform code using discovery data and Azure Verified Modules.

When to Use This Skill

Use this skill when the user asks to:

• Import existing Azure resources into Terraform
• Generate IaC from live Azure environments
• Handle any Azure resource type supported by AVM (and document justified non-AVM fallbacks)
• Recreate infrastructure from a subscription or resource group
• Map dependencies between discovered Azure resources
• Use AVM modules instead of handwritten azurerm_* resources

Prerequisites

• Azure CLI installed and authenticated (az login)
• Access to the target subscription or resource group
• Terraform CLI installed
• Network access to Terraform Registry and AVM index sources

Inputs

At least one of subscription-id, resource-group-name, or resource-id is required.

Step-by-Step Workflows

1) Collect Required Scope (Mandatory)

Request one of these scopes before running discovery commands:

• Subscription scope: <subscription-id>
• Resource group scope: <resource-group-name>
• Specific resources scope: one or more <resource-id> values

Scope handling rules:

• Treat Azure ARM resource IDs (for example /subscriptions/.../providers/...) as cloud resource identifiers, not local file system paths.
• Use resource IDs only with Azure CLI --ids arguments (for example az resource show --ids <resource-id>).
• Never pass resource IDs to file-reading commands (cat, ls, read_file, glob searches) unless the user explicitly says they are local file paths.
• If the user already provided one valid scope, do not ask for additional scope inputs unless required by a failing command.
• Do not ask follow-up questions that can be answered from already-provided scope values.

If scope is missing, ask for it explicitly and stop.

2) Authenticate and Set Context

Run only the commands required for the selected scope.

For subscription scope:

az login
az account set --subscription <subscription-id>
az account show --query "{subscriptionId:id, name:name, tenantId:tenantId}" -o json

Expected output: JSON object with subscriptionId, name, and tenantId.

For resource group or specific resource scope, az login is still required but az account set is optional if the active context is already correct.

When using specific resource scope, prefer direct --ids-based commands first and avoid extra discovery prompts for subscription or resource group unless needed for a concrete command.

3) Run Discovery Commands

Discover resources using the selected scopes. Ensure to fetch all necessary information for accurate Terraform generation.

# Subscription scope
az resource list --subscription <subscription-id> -o json

# Resource group scope
az resource list --resource-group <resource-group-name> -o json

# Specific resource scope
az resource show --ids <resource-id-1> <resource-id-2> ... -o json

Expected output: JSON object or array containing Azure resource metadata (id, type, name, location, tags, properties).

4) Resolve Dependencies Before Code Generation

Parse exported JSON and map:

• Parent-child relationships (for example: NIC -> Subnet -> VNet)
• Cross-resource references in properties
• Ordering for Terraform creation

IMPORTANT: Generate the following documentation and save it to a docs folder in the root of the project.

• exported-resources.json with all discovered resources and their metadata, including dependencies and references.
• EXPORTED-ARCHITECTURE.MD file with a human-readable architecture overview based on the discovered resources and their relationships.

5) Select Azure Verified Modules (Required)

Use the latest AVM version for each resource type.

Terraform Registry

• Search for "avm" + resource name
• Filter by "Partner" tag to find official AVM modules
• Example: Search "avm storage account" → filter by Partner

Official AVM Index

Note: The following links always point to the latest version of the CSV files on the main branch. As intended, this means the files may change over time. If you require a point-in-time version, consider using a specific release tag in the URL.

• Terraform Resource Modules: https://raw.githubusercontent.com/Azure/Azure-Verified-Modules/refs/heads/main/docs/static/module-indexes/TerraformResourceModules.csv
• Terraform Pattern Modules: https://raw.githubusercontent.com/Azure/Azure-Verified-Modules/refs/heads/main/docs/static/module-indexes/TerraformPatternModules.csv
• Terraform Utility Modules: https://raw.githubusercontent.com/Azure/Azure-Verified-Modules/refs/heads/main/docs/static/module-indexes/TerraformUtilityModules.csv

Individual Module information

Use the web tool or another suitable MCP method to get module information if not available locally in the .terraform folder.

Use AVM sources:

• Registry: https://registry.terraform.io/modules/Azure/<module>/azurerm/latest
• GitHub: https://github.com/Azure/terraform-azurerm-avm-res-<service>-<resource>

Prefer AVM modules over handwritten azurerm_* resources when an AVM module exists.

When fetching module information from GitHub repositories, the README.md file in the root of the repository typically contains all detailed information about the module, for example: https://raw.githubusercontent.com/Azure/terraform-azurerm-avm-res--/refs/heads/main/README.md

5a) Read the Module README Before Writing Any Code (Mandatory)

This step is not optional. Before writing a single line of HCL for a module, fetch and read the full README for that module. Do not rely on knowledge of the raw azurerm provider or prior experience with other AVM modules.

For each selected AVM module, fetch its README:

https://raw.githubusercontent.com/Azure/terraform-azurerm-avm-res-<service>-<resource>/refs/heads/main/README.md

Or if the module is already downloaded after terraform init:

cat .terraform/modules/<module_key>/README.md

From the README, extract and record before writing code:

1. Required Inputs — every input the module requires. Any child resource listed here (NICs, extensions, subnets, public IPs) is managed inside the module. Do not create standalone module blocks for those resources.
2. Optional Inputs — the exact Terraform variable names and their declared type. Do not assume they match the raw azurerm provider argument names or block shapes.
3. Usage examples — check what resource group identifier is used (parent_id vs resource_group_name), how child resources are expressed (inline map vs separate module), and what syntax each input expects.

Apply module rules as patterns, not assumptions

Use the lessons below as examples of the *type* of mismatch that often causes imports to fail. Do not assume these exact names apply to every AVM module. Always verify each selected module's README and variables.tf.

avm-res-compute-virtualmachine (any version)

• network_interfaces is a Required Input. NICs are owned by the VM module. Never create standalone avm-res-network-networkinterface modules alongside a VM module — define every NIC inline under network_interfaces.
• TrustedLaunch is expressed through the top-level booleans secure_boot_enabled = true and vtpm_enabled = true. The security_type argument exists only under os_disk for Confidential VM disk encryption and must not be used for TrustedLaunch.
• boot_diagnostics is a bool, not an object. Use boot_diagnostics = true; use the separate boot_diagnostics_storage_account_uri variable if a storage URI is needed.
• Extensions are managed inside the module via the extensions map. Do not create standalone extension resources.

avm-res-network-virtualnetwork (any version)

• This module is backed by the AzAPI provider, not azurerm. Use parent_id (the full resource group resource ID string) to specify the resource group, not resource_group_name.
• Every example in the README shows parent_id; none show resource_group_name.

Generalized takeaway for all AVM modules:

• Determine child resource ownership from Required Inputs before creating sibling modules.
• Determine accepted variable names and types from Optional Inputs and variables.tf.
• Determine identifier style and input shape from README usage examples.
• Do not infer argument names from raw azurerm_* resources.

6) Generate Terraform Files

Before Writing Import Blocks — Inspect Module Source (Mandatory)

After terraform init downloads the modules, inspect each module's source files to determine the exact Terraform resource addresses before writing any import {} blocks. Never write import addresses from memory.

Step A — Identify the provider and resource label

grep "^resource" .terraform/modules/<module_key>/main*.tf

This reveals whether the module uses azurerm_* or azapi_resource labels. For example, avm-res-network-virtualnetwork exposes azapi_resource "vnet", not azurerm_virtual_network "this".

Step B — Identify child modules and nested paths

grep "^module" .terraform/modules/<module_key>/main*.tf

If child resources are managed in a sub-module (subnets, extensions, etc.), the import address must include every intermediate module label:

module.<root_module_key>.module.<child_module_key>["<map_key>"].<resource_type>.<label>[<index>]

Step C — Check for count vs for_each

grep -n "count\|for_each" .terraform/modules/<module_key>/main*.tf

Any resource using count requires an index in the import address. When count = 1 (e.g., conditional Linux vs Windows selection), the address must end with [0]. Resources using for_each use string keys, not numeric indexes.

Known import address patterns (examples from lessons learned)

These are examples only. Use them as templates for reasoning, then derive the exact addresses from the downloaded source code for the modules in your current import.

Produce:

• providers.tf with azurerm provider and required version constraints
• main.tf with AVM module blocks and explicit dependencies
• variables.tf for environment-specific values
• outputs.tf for key IDs and endpoints
• terraform.tfvars.example with placeholder values

Diff Live Properties Against Module Defaults (Mandatory)

After writing the initial configuration, compare every non-zero property of each discovered live resource against the default value declared in the corresponding AVM module's variables.tf. Any property where the live value differs from the module default must be set explicitly in the Terraform configuration.

Pay particular attention to the following property categories, which are common sources of silent configuration drift:

• Timeout values (e.g., Public IP idle_timeout_in_minutes defaults to 4; live deployments often use 30)
• Network policy flags (e.g., subnet private_endpoint_network_policies defaults to "Enabled"; existing subnets often have "Disabled")
• SKU and allocation (e.g., Public IP sku, allocation_method)
• Availability zones (e.g., VM zone, Public IP zone)
• Redundancy and replication settings on storage and database resources

Retrieve full live properties with explicit az commands, for example:

az network public-ip show --ids <resource_id> --query "{idleTimeout:idleTimeoutInMinutes, sku:sku.name, zones:zones}" -o json
az network vnet subnet show --ids <resource_id> --query "{privateEndpointPolicies:privateEndpointNetworkPolicies, delegation:delegations}" -o json

Do not rely solely on az resource list output, which may omit nested or computed properties.

Pin module versions explicitly:

module "example" {
	source  = "Azure/<module>/azurerm"
	version = "<latest-compatible-version>"
}

7) Validate Generated Code

Run:

terraform init
terraform fmt -recursive
terraform validate
terraform plan

Expected output: no syntax errors, no validation errors, and a plan that matches discovered infrastructure intent.

Troubleshooting

Response Contract

When returning results, provide:

1. Scope used (subscription, resource group, or resource IDs)
2. Discovery files created
3. Resource types detected
4. AVM modules selected with versions
5. Terraform files generated or updated
6. Validation command results
7. Open gaps requiring user input (if any)

Execution Rules for the Agent

• Do not continue if scope is missing.
• Do not claim successful import without listing discovered files and validation output.
• Do not skip dependency mapping before generating Terraform.
• Prefer AVM modules first; justify each non-AVM fallback explicitly.
• Read the README for every AVM module before writing code. Required Inputs identify which child resources the module owns. Optional Inputs document exact variable names and types. Usage examples show provider-specific conventions (parent_id vs resource_group_name). Skipping the README is the single most common cause of code errors in AVM-based imports.
• Never assume NIC, extension, or public IP resources are standalone. For any AVM module, treat child resources as parent-owned unless the README explicitly indicates a separate module is required. Check Required Inputs before creating sibling modules.
• Never write import addresses from memory. After terraform init, grep the downloaded module source to discover the actual provider (azurerm vs azapi), resource labels, sub-module nesting, and count vs for_each usage before writing any import {} block.
• Never treat ARM resource IDs as file paths. Resource IDs belong in Azure CLI --ids arguments and API queries, not file IO tools. Only read local files when a real workspace path is provided.
• Minimize prompts when scope is already known. If subscription, resource group, or specific resource IDs are already provided, proceed with commands directly and only ask a follow-up when a command fails due to missing required context.
• Do not declare the import complete until terraform plan shows 0 destroys and 0 unwanted changes. Telemetry + create resources are acceptable. Any ~ update or - destroy on real infrastructure resources must be resolved.

References

• Azure Verified Modules index (Terraform)
• Terraform AVM Registry namespace