| name | thermo-nuclear-code-quality-review |
| description | Strict maintainability review for branch/PR diffs: thermonuclear/deep code quality, code-judo simplification, 1k-line growth, spaghetti conditionals, abstraction/type/boundary debt. Do not use for whole-repo audits, ponytail-only reviews, quick nits, security-only reviews, LGTM, or bug triage. |
Thermo-Nuclear Code Quality Review
Goal
Deliver a demanding maintainability review that pushes for structural simplification without changing intended behavior.
Success criteria
- Branch diff audited against the full rubric in
references/review-rubric.md.
- Review stays read-only: findings and sign-off only, no file edits through this skill.
- Context scoped to the diff's blast radius: no whole-repo scan, no diff-only tunnel vision.
- Findings cite concrete path/line or changed-hunk evidence, structural impact, and preferred remediation.
- Findings prioritized: structural regressions and missed simplification first.
- Approval withheld when presumptive blockers apply unless clearly justified.
- Few high-conviction comments over cosmetic nit floods.
Use when
- User wants thermo-nuclear / thermonuclear / deep maintainability review.
- PR needs harsh scrutiny on structure, file size, and abstraction—not just correctness.
When not to use
- Quick style pass or typo hunt.
- Security-only or incident triage without maintainability scope.
- Rubber-stamp "LGTM" requests.
Workflow
- Read
references/review-rubric.md completely before reviewing.
- Gather context per the scope contract below.
- For each meaningful change, run the primary review questions in the rubric.
- Flag presumptive blockers explicitly; propose preferred remedies from the rubric.
- Order output per output expectations in the rubric. If fixes are requested, return the review and recommended remediation path rather than editing files.
Scope contract
The diff is the seed, not the boundary. Never scan the whole repo.
- Seed (always):
git diff <base>...HEAD (triple-dot diffs from the merge-base; default base main), full contents of changed files, and wc -l on changed files for the 1k-line rule.
- Expand one hop at a time, only to answer a specific rubric question:
- Changed/new exported symbol or signature:
rg its callers; read only the call sites.
- Diff calls something unfamiliar: read that definition.
- Suspected duplicate of a canonical helper: search shared/util layers by name or keyword; read only the hits.
- Boundary/layer question: list the surrounding module and skim neighbor interfaces, not full bodies.
- Stop expanding when the question is answered, when a hop yields no new evidence, or at two hops from a changed symbol. If a concern truly needs repo-wide evidence, raise it as a question to the author with what you did verify; do not go scan for it.
Whole-repo health is a codebase audit, not this review.
Harness notes
- Default everywhere: gather scoped context per the scope contract, read
references/review-rubric.md, review inline in the parent session.
- Optional delegation: see
agents/subagent.md when the harness supports subagents (Cursor Task, pi sub-session, etc.).
Approval bar
Do not approve on behavior alone. Approve only when:
- no clear structural regression
- no obvious missed code-judo simplification
- no unjustified 1k+ file growth
- no obvious spaghetti from special-case branching
- no magic/wrapper/cast churn obscuring design
- no boundary leak or canonical-helper duplication
- no missed obvious decomposition
Otherwise leave actionable feedback and request restructuring.
Stop rules
Stop when high-priority findings are documented with clear remediation paths, or when the inspected scope has no material thermo-nuclear findings. Do not drown the author in low-value nits if structural issues remain.
Anti-patterns
- Mild suggestions for major structural debt.
- Approving because tests pass.
- Long lists of cosmetic notes while missing 1k-line or spaghetti regressions.
