Review checklist

This is the checklist to run down — by eye, on any diff, yours or an AI’s — after the tools in Verification Tools have run and come back clean. It fits on one printed page on purpose: a checklist nobody actually opens during a review is worthless. Every line links the entry that explains why, so this page stays a list, not a re-run of the course. The groups follow the same order as How to read Terraform you did not write, plus a final Tooling group covering how each of the other seven actually gets verified.

flowchart LR
  A["Versions &amp; providers"] --> T["validate/plan catch some<br/>(deprecation warnings, module+count errors)"]
  B["State &amp; backend"] --> P["Partial: remote_state hard-fails loudly,<br/>backend-locking drift is silent"]
  V["Variables"] --> M["Mostly human review<br/>(nullable/ephemeral behave silently)"]
  R["Resources &amp; lifecycle"] --> P2["Partial: for_each-on-unknown caught,<br/>prevent_destroy-on-wrong-resource is not"]
  Mo["Modules"] --> P
  Tg["Terragrunt live tree"] --> M
  C["CI &amp; policy"] --> P2

What actually catches each group, if anything

Versions & providers

• Is every required_providers constraint ~>, not >=? A floating constraint resolves to whatever the registry’s current newest release is on the day init runs. See the AI Bug Catalog, entry 1.
• Is .terraform.lock.hcl committed — not gitignored alongside .terraform/? See entry 4.
• Does terraform validate -json come back with zero warnings, not just zero errors? A deprecated argument (acl, managed_policy_arns) still validates and plans successfully. See entry 2.
• Does any module declare its own provider block? It validates fine alone and only fails once a caller needs count/for_each/depends_on. See entry 3.

State & backend

• Does the backend’s locking mechanism match what this Terraform/backend version actually supports — use_lockfile (S3, stable since 1.11) rather than the now-deprecated dynamodb_table? See the AI Bug Catalog, entry 5, and The State File.
• Does every terraform_remote_state data source point at a key/path that has actually been applied? There is no mock fallback — a missing state file hard-fails the whole plan. See entry 6.

Variables

• Does a nullable = false variable’s default ever silently stand in for an explicit null passed from a caller, with nothing in plan/apply output calling out the substitution? See the AI Bug Catalog, entry 7.
• Does any ephemeral variable carry a literal secret default? It still leaks into a saved plan’s show -json output even though it never reaches state. See entry 8.

Resources & lifecycle

• Is count used over any list that could be reordered, where for_each is the only safe choice? Removing a middle element shifts every later index. See the AI Bug Catalog, entry 10.
• Does lifecycle { prevent_destroy = true } sit on the resource that actually matters — checked directly, not assumed from a nearby resource’s error message? See entry 9, proven with a real plan -destroy that destroyed the unprotected database with zero resistance.
• Does any for_each derive its keys from another managed resource’s computed attribute? A real, current “Invalid for_each argument” rejection — see entry 11.
• Is null_resource used where terraform_data (1.4+) removes an entire extra provider dependency for equivalent behavior? See entry 12.
• Does any argument look copied from a similar-but-different resource type (force_destroy on google_compute_instance, real on google_storage_bucket but not here)? A real “Unsupported argument” rejection — see entry 13.

Modules

• Does every non-registry source pin a revision — ?ref=<tag-or-sha> for a git:: source, an explicit version for a registry module? Without it, Terraform tracks the default branch’s HEAD, silently. See the AI Bug Catalog, entry 14.
• Does any depends_on on a module call exist to “fix” an ordering complaint that’s actually a genuine reference cycle? No depends_on can suppress a real Error: Cycle. See entry 15.

Terragrunt live tree

• Does anything reference run-all/plan-all? Both are unknown commands as of the Terragrunt 1.1 CLI redesign — a real, immediate rejection. See the AI Bug Catalog, entry 16.
• Do two environments’ terragrunt.hcl files ever carry the identical account_id/region/subscription/project value in inputs? terragrunt hcl validate cannot catch a syntactically valid copy-paste. See entry 17, proven with a real byte-identical dev/prod diff.

CI & policy

• Does -auto-approve ever run against a fresh plan, rather than a saved, reviewed plan -out file? See the AI Bug Catalog, entry 18, proven with a real side-by-side apply.
• Is every third-party GitHub Action pinned to a full commit SHA, not a mutable tag like @master/@main? See entry 19.
• Does any Rego policy use pre-v1 deny[msg] { ... } syntax? It fails to even load under current OPA/conftest — see entry 20, proven with a real rego_parse_error.
• Does every .tflint.hcl rule reference something this small, IAM-focused tflint-ruleset-google (six rules total) actually offers — rather than assuming a resource-shape check other clouds’ larger rulesets have? See entry 21, proven with a real, immediate “Rule not found” rejection here.
• Is a real secret ever committed in a .tfvars file, or left in terraform.tfstate as plaintext despite an output marked sensitive? See entries 22–23.

Tooling

• Does terraform validate -json | jq '.diagnostics[]' get checked for warnings, not just a nonzero exit code? See Verification Tools.
• Does CI run a real terraform plan -detailed-exitcode and inspect the saved plan’s show -json output with a real policy tool (conftest, checkov), not just eyeball the plan text? See Verification Tools.
• Does terragrunt hcl validate/render run in CI before run --all plan, catching syntax errors before a 10-unit tree partially applies? See Verification Tools.

🏗️ A checklist narrows the search, it does not replace it

Every line above maps to one of the 23 patterns in the AI Bug Catalog — only a minority of which validate/plan/terragrunt reject outright. Run this list on every diff regardless of how clean those tools came back.

Why does the 'Versions & providers' group's coverage diagram say 'validate/plan catch some' while 'Variables' and 'Terragrunt live tree' say 'mostly human review'?

It's arbitrary — every group has identical tool coverage A deprecated argument or a module+count conflict is something validate/plan can detect from the schema/graph alone; whether nullable's default silently substituted for null, or whether two environments' terragrunt.hcl inputs genuinely differ, are semantic questions neither tool can check without knowing the reviewer's actual intent Variables are never validated by any tool Terragrunt files always fail hcl validate

A stack has a real database resource with no lifecycle block at all, sitting next to a disposable helper resource that carries lifecycle { prevent_destroy = true }. Which checklist group covers reviewing this pair, and what does it depend on to catch it?

Terragrunt live tree — it depends on per-environment inputs Resources & lifecycle — catching it depends on checking prevent_destroy against the resource that actually matters, since terraform plan itself accepts this configuration with zero complaint until a real plan -destroy is attempted Versions & providers — prevent_destroy is a provider-version concern Tooling — only checkov catches this

Why does the checklist ask 'is every Rego policy using v1 syntax' under CI & policy rather than assuming conftest would just warn about old syntax?

conftest always auto-upgrades old Rego syntax silently Proven directly: pre-v1 deny[msg] { ... } syntax produces a real rego_parse_error at LOAD time under current conftest — it's a hard failure to even start, not a warning, which can look like a CI/tooling problem rather than what it actually is Rego v1 and pre-v1 syntax are functionally identical This only matters for checkov, not conftest

Why does the checklist list 'does this .tflint.hcl rule actually exist in the real google ruleset' as its own line, given entry 21 shows tflint rejecting the invented rule immediately here?

It doesn't need to be — tflint always errors on every invalid rule name, for every ruleset Because whether an invented rule name is caught at all depends on the specific ruleset plugin — proven directly, this catalog's aws sibling shows the aws ruleset silently ignoring an invented rule, while the small google ruleset here rejects one outright; assuming either behavior universally is wrong This is only a theoretical concern, never observed tflint rejects unknown plugins entirely

Submit

Mark as complete