tf-plan-review（工作重点计划审查）

Terraform Plan Analyzer & Risk Assessor

Analyze terraform plan output and produce an AI-powered risk assessment of every infrastructure change — before you press apply.

This skill is STRICTLY READ-ONLY. It runs terraform plan and terraform validate to analyze changes, but it NEVER runs terraform apply, terraform destroy, terraform import, terraform taint, or any command that modifies infrastructure or state.

Activation

This skill activates when the user mentions:

• "terraform plan", "tf plan", "review plan", "plan review"
• "is this plan safe", "safe to apply", "risk assessment"
• "what will be destroyed", "what changes", "terraform changes"
• "terraform state", "state drift", "drift detection"
• "terraform validate", "validate config", "tf validate"
• "IAM changes", "security group changes", "infrastructure changes"
• "blast radius", "cascade effects", "dependencies"
• "tofu plan", "opentofu" (same workflow, different binary)

Example Prompts

1. "Review this terraform plan before I apply"
2. "What will be destroyed in this plan?"
3. "Is this plan safe to apply?"
4. "Show me the state drift"
5. "What IAM changes are in this plan?"
6. "Validate my terraform config in ~/infra/prod"
7. "Run a risk assessment on the terraform plan in /deployments/staging"
8. "What's the blast radius if I apply this plan?"

Permissions

permissions:
  exec: true          # Required to run terraform/tofu CLI
  read: true          # Read .tf files and plan output
  write: false        # NEVER writes — strictly read-only analysis
  network: true       # terraform plan needs provider API access

Terraform Change Types — What the Agent Must Know

Understanding Terraform change types is critical for accurate risk assessment:

Action Types (from plan JSON)

What Makes a Change Dangerous

Critical (🔴 CRITICAL):

• Any destroy/replace of: IAM roles/policies, security groups, KMS keys, secrets, databases (RDS, DynamoDB, Cloud SQL, Azure SQL), S3 buckets, DNS records, WAF rules, CloudTrail
• Any update to IAM policies, security group rules, encryption settings
• These changes can cause data loss, security breaches, or service outages

Dangerous (🟠 DANGEROUS):

• Destroy/replace of: EC2 instances, load balancers, ECS/EKS clusters, VPCs, subnets, NAT gateways, Lambda functions, API gateways
• These changes cause downtime and may require manual intervention to recover

Moderate (🟡 MODERATE):

• Updates to: autoscaling policies, monitoring/alerting rules, launch templates
• Creates of: security-sensitive resources (new IAM roles, new security groups)
• Changes that affect capacity or observability but not data integrity

Safe (🟢 SAFE):

• Tag-only updates
• Creating new non-sensitive resources
• No-op / read operations

Replace is Especially Dangerous

When Terraform says it must "replace" a resource, it means:

1. Delete the existing resource (irreversible)
2. Create a new one with the new configuration

This is triggered when an immutable attribute changes (e.g., changing RDS engine_version, EC2 ami, changing a subnet's AZ). The agent should always flag replaces prominently because:

• The old resource (and its data) is destroyed
• There will be a gap between destroy and create (downtime)
• Dependent resources may break during the transition

Agent Workflow

Follow this sequence exactly based on user intent:

For Plan Analysis ("review this plan", "is it safe", "what changes")

Step 1: Run Plan Analysis

bash <skill_dir>/scripts/tf-plan-review.sh plan <directory>

If no directory specified, use the current working directory.

The script outputs:

• stdout: Structured JSON with all resource changes, risk classifications, and summary
• stderr: Beautiful Markdown risk report

Step 2: Interpret the JSON

Parse the JSON output. Key fields:

{
  "overall_risk": "🔴 CRITICAL | 🔴 HIGH | 🟡 MODERATE | 🟢 LOW",
  "summary": {
    "create": 5,
    "update": 3,
    "destroy": 1,
    "replace": 0
  },
  "risk_breakdown": {
    "critical": 1,
    "dangerous": 0,
    "moderate": 2,
    "safe": 5
  },
  "resources": [
    {
      "address": "aws_iam_role.admin",
      "action": "delete",
      "risk": "🔴 CRITICAL"
    }
  ]
}

Step 3: Present the Risk Assessment

Show the Markdown report from stderr. Then add your own AI analysis:

1. Lead with the overall risk level — make it viscerally clear
2. Highlight destroys and critical changes first — these are what kill production
3. Explain WHY each critical change is dangerous in plain English
4. Assess blast radius — what other resources depend on the destroyed ones?
5. Present the pre-apply checklist — what should the human verify?
6. Give a clear recommendation: "Safe to apply" / "Review needed" / "DO NOT APPLY without ___"

Tone guidance for critical plans:

• Don't be polite about danger. If a plan destroys a production database, say so bluntly.
• "This plan will permanently delete your RDS instance prod-db. All data will be lost. Do you have a backup?"
• Make the "oh shit" moment impossible to miss.

For State Inspection ("show me state", "what's managed", "state drift")

bash <skill_dir>/scripts/tf-plan-review.sh state "<filter>" <directory>

The filter is optional — it greps resource addresses. Examples:

• bash <skill_dir>/scripts/tf-plan-review.sh state "iam" . → all IAM resources
• bash <skill_dir>/scripts/tf-plan-review.sh state "aws_instance" . → all EC2 instances
• bash <skill_dir>/scripts/tf-plan-review.sh state "" . → all resources

For Validation ("validate config", "check syntax")

bash <skill_dir>/scripts/tf-plan-review.sh validate <directory>

Reports configuration errors and warnings without running a plan.

Environment Variables

The script auto-detects terraform first, then tofu. Set TF_BINARY=tofu to force OpenTofu.

Error Handling

Safety — CRITICAL RULES

1. NEVER run terraform apply — not even with -auto-approve, not even with -target, not even "just this one resource". NEVER.
2. NEVER run terraform destroy — not under any circumstances.
3. NEVER run terraform import — this modifies state.
4. NEVER run terraform taint or terraform untaint — these modify state.
5. NEVER run terraform state mv, terraform state rm, or terraform state push — these modify state.
6. Never expose cloud credentials — if they appear in plan output, redact them.
7. Handle sensitive values — Terraform marks values as (sensitive). Never try to reveal them.
8. Never cache or store plan output — plans can contain secrets in resource attributes.
9. The ONLY terraform commands this skill runs are: plan, show, state list, state show, validate, init, providers.

If the user asks you to apply a plan, respond:

"I can analyze and assess Terraform plans, but I cannot apply them. Applying infrastructure changes requires human review and explicit execution. Based on my analysis, here's what you should verify before running terraform apply..."

Common Patterns & Agent Tips

"Is this plan safe to apply?"

Run the plan analysis. If overall_risk is 🟢 LOW:

"This plan looks safe. It creates X new resources with no destroys or security changes. The pre-apply checklist is straightforward."

If overall_risk is 🔴 CRITICAL:

"⚠️ This plan has CRITICAL risk. [Explain specific dangers]. I strongly recommend review by another team member before applying."

"What will be destroyed?"

Run plan, then filter for action == "delete" or action == "replace". Present each with:

• Resource address
• Resource type
• Why it matters (is it stateful? does it have data?)
• What depends on it

"What IAM changes are in this plan?"

Run plan, then filter resources matching IAM patterns. For each:

• What permission is changing
• Is it adding or removing access
• Is it overly permissive (e.g., Action: *)

"Show me the blast radius"

Run plan, identify all destroys/replaces, then explain:

• What other resources reference the destroyed ones
• What will break when the resource is gone
• Whether Terraform will auto-fix the dependencies or if manual intervention is needed

Discord v2 Delivery Mode (OpenClaw v2026.2.14+)

When the conversation is happening in a Discord channel:

• Send a compact first summary (overall risk, destroy count, critical resources), then ask if the user wants the full report.
• Keep the first response under ~1200 characters and avoid large Markdown tables in the first message.
• If Discord components are available, include quick actions:

- Show Critical Changes - Show Destroyed Resources - Show Pre-Apply Checklist

• If components are not available, provide the same follow-ups as a numbered list.
• Prefer short follow-up chunks (<=15 lines per message) for large plans.

Sensitive Data Handling

Terraform plan JSON may contain sensitive values. The script does NOT extract resource attribute values — it only extracts resource addresses, types, and actions. However, when presenting results:

• Never show attribute values marked (sensitive) by Terraform
• Never show provider credentials or backend configuration secrets
• If a user asks "what value is changing?", explain that you can see the change type but sensitive values are redacted by Terraform for security
• Never store or cache plan output files

Powered by Anvil AI 🔍