Tf Best Practices

tf-best-practices — Generated-IaC posture rules + read-only policy gate

A shared authoring guide and verdict producer, not a workflow. It answers two questions for a phase that generates AWS Terraform:

1. Before writing — “what security posture must the generated terraform/ follow?” (the posture rules + the baseline.tf account-hardening spec)
2. After writing — “does the generated terraform/ pass policy?” (a deterministic, read-only verdict + a machine-readable report)

Routing — load the part that matches your context

This skill is entered at two touchpoints in the caller’s Generate flow, with the caller’s own terraform-authoring work in between. The caller states which touchpoint it is at when it loads this skill, and reads the corresponding part:

Everything this skill states is source-cloud-agnostic (pure AWS Terraform). Any GCP/Heroku detection or artifact reading is the caller’s job; where a rule needs a caller-known fact (e.g. declared compliance frameworks), the caller passes it as a caller-context signal — see references/security-posture-rules.md § Caller-context signals.

Boundary (read this first)

This unit is a verdict producer, never a mutator. Its entire write surface is the JSON verdict it is asked to emit. Specifically it MUST NOT:

• edit, format, or rewrite any .tf file (the caller owns remediation),
• read or write .phase-status.json or any run-state file (interpreter-owned),
• decide whether a phase may complete, or prompt the user (caller policy).

The caller (a migration skill’s Generate phase) owns: the fix-and-retry loop that edits the .tf it generated, terraform fmt auto-apply, the retry/skip/abort prompt, the Phase Completion gate, and every .phase-status.json write. See the consuming skill’s generate phase for how the verdict feeds those decisions.

Consumers: gcp-to-aws (prose Generate) and heroku-to-aws (DSL Generate). The contract is source-agnostic; each caller wires the two touchpoints in its own Generate idiom — gcp-to-aws as prose steps, heroku-to-aws as a fragment step plus a fail-closed _postconditions assert enforced by the interpreter.

Part 1 — Authoring posture (load before writing terraform/)

Emit generated Terraform that satisfies the posture in references/security-posture-rules.md (opens in a new tab).

These are the “what good AWS Terraform looks like” rules. Following them makes the Part 2 gate pass by construction. This unit does not read the caller’s artifacts — it consumes only caller-context signals the caller passes in.

Scope. security-posture-rules.md covers, in three tiers:

• Gate-enforced (Part 2 verifies statically): ALB TLS, no-public-database, RDS + ElastiCache encryption-at-rest, no-public-DB-port ingress, no-public admin/datastore-port ingress, no-wildcard-IAM.
• Authoring-only (not gate-checkable, still required): deletion_protection, master-password-via-Secrets-Manager, S3 hardening, Fargate/EKS/ECR settings, private-subnet placement, backups, baseline monitoring.
• Compliance-conditional (emitted when the caller declares soc2/pci/hipaa/fedramp): VPC flow logs, S3 access logging, secret rotation, customer-managed KMS.

Still the caller’s own generation concern (candidates to migrate here later): the account-hardening baseline.tf layer (CloudTrail, GuardDuty, Config, Security Hub).

Part 2 — Policy gate (run after writing terraform/)

Run the read-only checker against the generated directory. Resolve the script path relative to the plugin root ($PLUGIN_ROOT/skills/tf-best-practices/scripts/...), the same convention the plugin uses for its other scripts:

python3 "$PLUGIN_ROOT/skills/tf-best-practices/scripts/validate-terraform-policy.py" "$TERRAFORM_DIR" --json "$VERDICT_PATH"

• $TERRAFORM_DIR — required, caller-supplied: the generated terraform/ directory (e.g. $MIGRATION_DIR/terraform). This skill never defaults or discovers it — the caller always passes the path it wrote Terraform to.
• --json $VERDICT_PATH — optional; writes a machine-readable verdict the caller can merge into its own validation-report.json.

The policy check is one stage of a larger validation flow (fmt → init → validate → policy). The full protocol — including offline-fallback behavior and how the policy verdict maps into a validation-report.json — is documented in references/terraform-validation.md (opens in a new tab). That protocol is descriptive: the caller owns the fmt/init/validate execution, the fix-and-retry loop, and the report write; this unit contributes only the read-only policy stage + verdict shape.

Exit codes → caller action

Verdict shape (--json)

{
  "check": "policy",
  "policy_status": "POLICY_OK | POLICY_FAIL",
  "violations": [
    {
      "check": "policy",
      "rule": "alb_https_listener | alb_http_redirect | no_tf_files",
      "file": "compute.tf",
      "line": 7,
      "severity": "error"

Each violations[] entry is actionable evidence — file + line + fix_hint tell the caller exactly what to edit. The caller applies the edit; this unit only reports.

Policy rules enforced today

Every rule is fail-open on ambiguity — it fires only on unambiguous, in-block literal evidence, so a valid stack is never falsely blocked (a POLICY_FAIL is a hard completion gate for the caller, so a false positive would block a real migration).

Internet-facing ALB TLS posture (an ALB is internet-facing when internal is absent, false, or variable-driven — fail-safe):

• alb_https_listener — must have an HTTPS listener on 443 with certificate_arn and a forward action.
• alb_http_redirect — an HTTP :80 listener must redirect to HTTPS, never forward to targets. Internal ALBs (internal = true) are exempt.

Elastic Beanstalk ALBs are invisible to these rules. The ALB rules inspect standalone aws_lb_listener blocks. An EB LoadBalanced environment provisions its ALB from aws_elastic_beanstalk_environment setting blocks, which the static checker does not read — so a pure-EB design passes the ALB rules vacuously (no listener to inspect). EB listener/TLS posture is therefore authoring-only, not gate-enforced. (Fixtures good-heroku-eb-only and good-heroku-eb-singleinstance document this; good-heroku-eb-loadbalanced carries a standalone ALB so the listener rules are exercised on real blocks.)

Managed database exposure & encryption (aws_db_instance, aws_rds_cluster):

• rds_not_public — must not set publicly_accessible = true (absent/variable → fail-open).
• rds_encryption_at_rest — must set storage_encrypted = true; missing or literal false fires (RDS defaults to unencrypted), variable-driven fails open. S3 is not checked (default SSE-S3 since Jan 2023).

ElastiCache encryption (aws_elasticache_replication_group, Redis aws_elasticache_cluster):

• elasticache_encryption_at_rest — a replication group must set at_rest_encryption_enabled = true; missing or literal false fires, variable-driven fails open.
• elasticache_cluster_encryption — a Redis-engine aws_elasticache_cluster (single-node: engine = "redis", no replication_group_id) must set BOTH at_rest_encryption_enabled = true and transit_encryption_enabled = true; missing or literal false on either fires, variable-driven fails open. engine = "memcached" clusters (and variable-driven/absent engine) are exempt — Memcached does not support these attributes.

Security group ingress:

• db_sg_no_public_ingress — an inline aws_security_group ingress covering 5432/3306 must not allow 0.0.0.0/0 or ::/0.
• sg_no_public_admin_ingress — an inline ingress must not open a curated never-public admin/datastore port (22, 3389, 6379, 11211, 27017, 9200/9300, 5601) to 0.0.0.0/0 or ::/0. Web (80/443) and app/game ports are not flagged; DB ports are handled by the rule above. Both check cidr_blocks and ipv6_cidr_blocks independently, so a benign IPv4 list does not mask an open IPv6 one. Both: separate aws_security_group_rule / aws_vpc_security_group_ingress_rule resources fail open (not correlated).

IAM least-privilege (aws_iam_policy, aws_iam_role_policy, aws_iam_group_policy, aws_iam_user_policy):

• no_wildcard_iam — an Allow statement must not use Action/Resource "*". The one narrow exception is an isolated elasticbeanstalk:CreateStorageLocation statement with Resource = "*" because AWS does not support resource-level permissions for that action. aws_iam_policy_document data sources and assume-role trust policies fail open.

The checker is a zero-dependency static HCL reader (no terraform init, no provider download) — it runs even when the registry is unreachable. It uses brace-depth matching for nested blocks, so a valid HTTPS listener written with a nested forward { ... } block is not a false failure.

Fixtures (also the checker’s regression suite)

fixtures/terraform-policy/ holds intentionally-shaped Terraform used by scripts/test_validate_terraform_policy.py:

• bad-http-forward/ — internet-facing ALB that forwards plaintext HTTP → MUST POLICY_FAIL.
• internal-alb-only/ — internal ALB on HTTP → MUST POLICY_OK (HTTP allowed internally).
• good-https-redirect/ — the correct pattern → POLICY_OK.

These are deliberately non-compliant test data (never deployed). They are excluded from the repo-wide checkov scan via .checkov.yaml skip-path; do not “harden” them — doing so breaks the tests that assert the failure paths.

Verification

# from skills/tf-best-practices/
uv run --python 3.12 --with pytest python -m pytest scripts/test_validate_terraform_policy.py -q