Require plan-destroy review before any terraform destroy command

POWER.md

@@ -25,6 +25,8 @@ Every Terraform/OpenTofu response must include:
 
 Never recommend direct production apply without a reviewed plan artifact and approval.
 
+Never run `terraform destroy` (targeted or full) without first running `terraform plan -destroy` and showing the user every resource that will be deleted — including implicit dependents pulled in via locals or `for_each`. Get explicit confirmation before proceeding. Never use `-auto-approve` on destroy.
+
 ## Workflow
 
 1. **Capture execution context** — runtime+version, provider(s), backend, execution path, environment criticality.
@@ -42,6 +44,7 @@ Never recommend direct production apply without a reviewed plan artifact and app
 | **Identity churn** | Resource addresses shift after refactor, `count` index churn, missing `moved` blocks | [Code Patterns: count vs for_each](skills/terraform-skill/references/code-patterns.md#count-vs-for_each-deep-dive), [Code Patterns: moved blocks](skills/terraform-skill/references/code-patterns.md#moved-blocks-terraform-11), [Code Patterns: LLM mistakes](skills/terraform-skill/references/code-patterns.md#llm-mistake-checklist--code-patterns) |
 | **Secret exposure** | Secrets in defaults, state, logs, CI artifacts | [Security & Compliance](skills/terraform-skill/references/security-compliance.md), [Code Patterns: write-only](skills/terraform-skill/references/code-patterns.md#write-only-arguments-terraform-111), [State Management](skills/terraform-skill/references/state-management.md) |
 | **Blast radius** | Oversized stacks, shared prod/non-prod state, unsafe applies | [State Management](skills/terraform-skill/references/state-management.md), [Module Patterns](skills/terraform-skill/references/module-patterns.md) |
+| **Destroy cascade** | Targeted destroy deletes more than expected; locals referencing a targeted resource make all `for_each` consumers implicit dependents | Response Contract: plan-destroy first; [State Management: Safe Destroy](skills/terraform-skill/references/state-management.md#safe-destroy-protocol) |
 | **CI drift** | Local plan ≠ CI plan, apply without reviewed artifact, unpinned versions | [CI/CD Workflows](skills/terraform-skill/references/ci-cd-workflows.md), [Code Patterns: versions](skills/terraform-skill/references/code-patterns.md#version-management) |
 | **Compliance gaps** | Missing policy stage, no approval model, no evidence retention | [Security & Compliance](skills/terraform-skill/references/security-compliance.md), [CI/CD Workflows](skills/terraform-skill/references/ci-cd-workflows.md) |
 | **Testing blind spots** | Plan-only validation of computed values, set-type indexing, mock/real confusion | [Testing Frameworks](skills/terraform-skill/references/testing-frameworks.md) |

skills/terraform-skill/SKILL.md

@@ -23,6 +23,8 @@ Every Terraform/OpenTofu response must include:
 
 Never recommend direct production apply without a reviewed plan artifact and approval.
 
+Never run `terraform destroy` (targeted or full) without first running `terraform plan -destroy` and showing the user every resource that will be deleted — including implicit dependents pulled in via locals or `for_each`. Get explicit confirmation before proceeding. Never use `-auto-approve` on destroy.
+
 ## Workflow
 
 1. **Capture execution context** — runtime+version, provider(s), backend, execution path, environment criticality.
@@ -40,6 +42,7 @@ Never recommend direct production apply without a reviewed plan artifact and app
 | **Identity churn** | Resource addresses shift after refactor, `count` index churn, missing `moved` blocks | [Code Patterns: count vs for_each](references/code-patterns.md#count-vs-for_each-deep-dive), [Code Patterns: moved blocks](references/code-patterns.md#moved-blocks-terraform-11), [Code Patterns: LLM mistakes](references/code-patterns.md#llm-mistake-checklist--code-patterns) |
 | **Secret exposure** | Secrets in defaults, state, logs, CI artifacts | [Security & Compliance](references/security-compliance.md), [Code Patterns: write-only](references/code-patterns.md#write-only-arguments-terraform-111), [State Management](references/state-management.md) |
 | **Blast radius** | Oversized stacks, shared prod/non-prod state, unsafe applies | [State Management](references/state-management.md), [Module Patterns](references/module-patterns.md) |
+| **Destroy cascade** | Targeted destroy deletes more than expected; locals referencing a targeted resource make all `for_each` consumers implicit dependents | Response Contract: plan-destroy first; [State Management: Safe Destroy](references/state-management.md#safe-destroy-protocol) |
 | **CI drift** | Local plan ≠ CI plan, apply without reviewed artifact, unpinned versions | [CI/CD Workflows](references/ci-cd-workflows.md), [Code Patterns: versions](references/code-patterns.md#version-management) |
 | **Compliance gaps** | Missing policy stage, no approval model, no evidence retention | [Security & Compliance](references/security-compliance.md), [CI/CD Workflows](references/ci-cd-workflows.md) |
 | **Testing blind spots** | Plan-only validation of computed values, set-type indexing, mock/real confusion | [Testing Frameworks](references/testing-frameworks.md) |

skills/terraform-skill/references/state-management.md

@@ -16,6 +16,7 @@ This document provides detailed guidance on state management, from remote backen
 5. [Multi-Team State Isolation](#multi-team-state-isolation)
 6. [State Recovery & Troubleshooting](#state-recovery--troubleshooting)
 7. [State Best Practices Summary](#state-best-practices-summary)
+   - [Safe Destroy Protocol](#safe-destroy-protocol)
 
 ---
 
@@ -1792,6 +1793,16 @@ rm .terraform/terraform.tfstate.lock.info  # ❌ Dangerous
 - High blast radius
 ✅ **Instead:** Split by component/team/environment
 
+### Safe Destroy Protocol
+
+A targeted destroy can cascade far beyond its targets via implicit dependencies. Always follow this sequence:
+
+1. Run `terraform plan -destroy [-target=...]` — never skip straight to destroy
+2. Read **every** resource under "will be destroyed", not just the targeted ones
+3. ⚠️ Watch for `for_each` resources fed by a local that references a targeted resource — all instances become implicit dependents (e.g. targeting an `aws_eip` that is referenced in a `locals {}` block used by a `for_each` record resource will destroy **all** those records)
+4. Get explicit user confirmation of the full list before proceeding
+5. ❌ Never use `-auto-approve` on destroy in production
+
 ### LLM Mistake Checklist — State Management
 
 Common model mistakes to correct before returning state-related recommendations:
@@ -1804,6 +1815,7 @@ Common model mistakes to correct before returning state-related recommendations:
 - mixes prod and non-prod in the same backend key
 - recommends workspace-only isolation as a substitute for backend-level IAM separation
 - writes DynamoDB-lock configuration on Terraform 1.10+ instead of using `use_lockfile = true` on the S3 backend
+- runs `terraform destroy -auto-approve` or skips `plan -destroy` before a targeted destroy — missing implicit dependents pulled in via shared locals
 - reads via `terraform_remote_state` within a single team's stack instead of using module outputs (see [module-patterns.md](module-patterns.md#3-use-terraform_remote_state-sparingly--only-at-true-ownership-boundaries))
 - omits the rollback/recovery note for destructive state operations