Code Review Auditor

Code Review Auditor

Use this skill when the user asks for a code review, architecture review, security review, refactoring review, design-pattern assessment, hotspot analysis, or a planned fix/refactor workflow.

Non-Negotiable Constraints

• Analysis modes never edit source code, tests, configuration, migrations, lockfiles, generated files, or infrastructure files.
• Refactoring and fixes must be planned before implementation. Create or update review//refactoring-plan.md first, then wait for explicit user approval unless the user has already asked to implement that plan.
• Avoid overengineering. Recommend a pattern, abstraction, split, layer, or framework only when it reduces current or near-term complexity and is justified by evidence in the code.
• Every execution must end by creating a local review folder in the project being reviewed:

review/YYYY-MM-DD_HH-mm-ss/

Use the local timezone for the timestamp. Include seconds. If two runs collide, append -NN while preserving the timestamp.

Supported Modes

Choose the narrowest mode that satisfies the user request. If the user does not specify a mode, use complete for broad review requests and diff when the request clearly focuses on changed code.

• complete: full project review across all categories.
• diff or changed: review only changed files and relevant call sites.
• security: vulnerabilities, insecure defaults, secrets, authz/authn, injection, SSRF, deserialization, path traversal, dependency risk.
• architecture: module boundaries, layering, SOLID, coupling, cohesion, domain leakage, framework misuse.
• smells: maintainability problems and local code-quality issues.
• patterns: justified opportunities for Strategy, Factory, Adapter, Decorator, Chain of Responsibility, State, Specification, Repository, Unit of Work, Observer, or other patterns already aligned with the project.
• performance: hot paths, query behavior, memory, concurrency, IO, caching, serialization, batch size, retries, backpressure.
• tests: missing, weak, brittle, slow, flaky, low-signal, or over-mocked tests.
• hotspots: prioritize files by churn, complexity, ownership, risk, dependency fan-in/fan-out, and production criticality.
• explain: explain findings, risks, and tradeoffs without changing code.
• fix: propose and, only after approval, implement narrowly scoped fixes.
• refactor: propose and, only after approval, implement behavior-preserving refactors.
• challenge: critique an existing implementation or proposal, looking for hidden failure modes and unjustified complexity.

Read workflows/modes.md when mode selection or mode-specific deliverables matter.

Required Review Folder

At the end of every execution, write the review package using templates/output-structure.md. At minimum include:

• summary.md
• findings.md or category files such as security.md, architecture.md, bugs.md, code-smells.md, patterns.md, performance.md, testing.md, and observability.md
• hotspots.md
• metrics/score.md
• metadata.json
• refactoring-plan.md when fixes, refactors, or material design changes are proposed

Use scripts/create_review_run.py to create the timestamped folder and seed the required files when helpful.

Finding Contract

Each finding must include:

• stable ID within the execution
• file and line range
• category
• title and description
• evidence from the code
• impact
• severity
• confidence
• effort
• priority
• recommendation
• suggested example or refactoring sketch
• possible false-positive note

Use the models in rules/scoring.md and the finding template in templates/finding.md.

Review Method

1. Identify the stack, framework, package manager, test tools, database layer, messaging systems, and deployment/infrastructure hints.
2. Read the changed files or representative architecture before judging. Prefer repository-local conventions over generic preferences.
3. Check category-specific rules from rules/categories.md, then read only the relevant stack guide from stacks/.
4. Verify each important finding against actual code paths, tests, configuration, or runtime contracts when possible.
5. De-duplicate findings. Prefer one root-cause finding over many symptoms.
6. Record false-positive possibilities instead of overstating certainty.
7. Assign severity, confidence, effort, priority, and refactorability score consistently.
8. Write all review artifacts before responding to the user.

Stack Guides

Read only the relevant guides:

• Java/Spring: stacks/java-spring.md
• Node/TypeScript: stacks/node-typescript.md
• Kotlin: stacks/kotlin.md
• Swift: stacks/swift.md
• React: stacks/react.md
• Databases: stacks/database.md
• Kafka: stacks/kafka.md

Output Tone

Be direct and evidence-led. Findings are not style opinions. Call out what can break, where, why it matters, and the smallest responsible improvement.