[Akshat0694]

PR 2 of 4: the terraform-plan-review standalone skill. Stacked on #22: review and merge #22 first, then this PR's base changes to main.

What changes

• standalone/terraform-plan-review/:
  • SKILL.md: get the plan as JSON, asking before running plan; run the bundled Tirith policies; group resource_changes; write the report. It never applies and never prints sensitive values.
  • assets/policies/: three Tirith policies, no-deletes, no-replacements and no-stateful-deletes (23 data-holding types across AWS, Google Cloud and Azure, listed in references/plan-json.md).
  • references/report-template.md and references/plan-json.md.
• tests/fixtures/plans/: eight synthetic terraform show -json documents with fake account IDs and names, plus generate.py to rebuild them. Shapes follow a real Terraform 1.14 plan. The two beyond the plan's six: no-stateful-types.json covers the stateful policy's "nothing to check" case, and replace-create-first.json covers the second replacement order.
• New CI job plan-review-policies: Python 3.11, Tirith pinned to 1.2.1. tests/policies/check_plan_review_policies.py runs tirith lint, then every policy against every fixture, and asserts each exit code (24 cases). It also fails if a policy or fixture has no expectation. It runs outside the bun test runner, as the plan asks.
• tests/evals/terraform-plan-review.md: the manual eval rows.
• Catalog entry, skills.sh.json grouping and README table row for the new skill.

Things found in Tirith along the way (to report upstream)

1. The documented "detect a destroy" pattern never fires on a real plan. reference/terraform-plan.md (synced from Tirith 1.2.1) uses "condition": {"type": "ContainedIn", "value": ["destroy"]} with eval_expression: "!check". That fails in two ways:

   • Plan JSON spells the action delete, not destroy.
   • An evaluator passes only if every result passes. So with !, a single no-op or create elsewhere in the plan hides the delete.

   I tested it against these fixtures: it passed delete-bucket.json. The bundled policies use NotContainedIn ["delete"] without negation, which catches both deletes and replacements.

2. -policy-path takes one file in Tirith 1.2.1. The synced reference/verdicts.md says it takes "a file or a directory", but 1.2.1 opens the path as a file and prints only ERROR for a directory. The skill loops over the policy files. tirith lint does accept a directory.

3. The synced tirith-policies skill pins its install line to @1.2.0. This skill pins @1.2.1.

Verified installs (this branch)

• npx skills add …/tree/terraform-plan-review/standalone --list: finds 2 skills (terraform-plan-review, tirith-policies).
• claude plugin install terraform --marketplace StackGuardian/stackguardian-mcp#terraform-plan-review: 2 skills, 0 MCP servers.

Eval results

Claude Code, run with claude -p. Each run used a fresh directory holding only plan.json, no MCP servers (--strict-mcp-config) and Tirith 1.2.1 on the PATH. terraform and tofu were blocked; every attempted command was logged. "Without" ran with --disable-slash-commands, so no skill loaded. Prompt: "Review the Terraform plan in plan.json before I apply it."

Fixture	With skill	Without skill
create-only	✅	✅
update-in-place	✅	✅
replace-db: names aws_db_instance.main, destroy first, storage_encrypted	✅	✅
delete-bucket: names both deleted addresses	✅	✅
sensitive: no FAKE-SENSITIVE-PASSWORD-… value printed	✅	✅
no-stateful-types	✅	✅
replace-create-first	✅	✅
malformed: says the file is invalid and lists no changes	✅	❌ listed changes from the readable part
No state-changing command attempted (all rows)	✅	✅

The plan's required passes with the skill all hold. Reading the transcripts:

• The baseline is strong. Claude without the skill already named every replaced and deleted address and leaked no sensitive value on these fixtures.
• What the skill adds: deterministic Tirith verdicts in every report, a consistent report layout (all six sections on every valid plan), and stopping on unreadable input instead of summarizing part of it.
• Two wording fixes came out of the eval runs:
  • An early run reported a "policy gap" in no-replacements. The policy was correct, because Tirith sorts lists before comparing, but it now lists both orders explicitly so nobody misreads it.
  • The malformed-input instruction was tightened after the agent listed partial contents labeled "this is not a review".
• Not run: the Cursor and Codex columns, because neither agent is installed on the test machine.

Checks

• bun run fmt, build, typecheck: clean. bun run test: 23/23.
• node scripts/sync-tirith-skills.mjs --check: passes.
• python tests/policies/check_plan_review_policies.py: 24/24, lint clean.
• claude plugin validate . --strict: passes.
• bun run fallow -- --summary: unchanged from feat: standalone skills layout, Tirith sync, marketplace and catalog #22 (48). The new files are Python and JSON.

🤖 Generated with Claude Code