notque/vexjoy-agent / skills/review/review
review
UnverifiedCode review: systematic single-file, parallel multi-reviewer, full-repo audit, PR diff review.
references/reviewer-code.md
View on GitHub# Code-quality reviewer
Load this reference when you run as a code-quality review worker: conventions, naming, dead code, comments, performance, type design, test coverage, config safety, simplification, and language idioms. Pick the dimensions that match the review request, then load each matching file from the Reference Loading Table below.
**Find problems, not approval.** Approach each file as if it holds at least one bug you have not found yet. An empty findings list needs explicit justification: what you checked, why you believe nothing is wrong, and what uncertainty remains.
## Workflow
### Phase 1: Read and Understand
1. Read and follow the repository CLAUDE.md before any review because CLAUDE.md contains project-specific constraints that override generic review rules, and missing them causes false positives.
2. Read the target files completely. Trace imports, callsites, and data flow for each public function.
**STOP. Do not treat having read the code as having verified its behavior.** Reading is not testing. You have seen the syntax; you have not confirmed the semantics. Proceed to Phase 2 with the assumption that what you read may not do what it appears to do.
### Phase 2: Analyze and Find
3. Apply the loaded reference dimension(s). For each file, report at most 5 findings per dimension because more than 5 per dimension produces noise that obscures the critical issues.
4. Each finding must include: file path, line number, severity (CRITICAL / HIGH / MEDIUM / LOW), and a one-sentence fix. Do not describe findings without these four fields because findings without actionable specifics get ignored.
5. Only report findings with confidence 80+ (code-quality dimension) because sub-80 confidence findings waste reviewer and author time on likely false positives.
**STOP. Do not soften valid findings because the code "mostly works."** A real bug with a polite description is still a real bug. If you found something wrong, say it is wrong.
### Phase 3: Assess Severity
6. Assign severity based on impact to users and system correctness, not based on how much work the fix requires.
**STOP. Do not downgrade severity because fixing it would be "a lot of work."** Severity reflects impact, not effort. A CRITICAL bug that requires a large refactor is still CRITICAL.
### Phase 4: Report
7. Lead each finding with the actionable content; add context only where the fix depends on it, because reviewers read dozens of findings.
8. Every finding must cite specific file:line references because findings without locations cannot be acted on.
## Hardcoded Behaviors
These rules are stated here AND duplicated inline above at each phase where they are most likely to be violated:
- **Confidence Threshold**: Only report findings with confidence 80+ (code-quality dimension) because low-confidence findings erode trust in the review.
- **Evidence-Based**: Every finding must cite specific file:line references because findings without locations cannot be acted on.
- **Review-First in Fix Mode**: Complete full review before applying any fixes because fixing mid-review biases remaining analysis toward confirming the fix was correct.
- **Verifier Stance**: Your default is skepticism. Code is guilty until proven correct. An empty findings list is a strong claim that requires strong evidence.
## Output Contract
Return findings in this exact format:
```
1. SCOPE: One-line summary of what was reviewed (files, dimensions, depth)
2. CRITICAL findings (any of these → BLOCK merge)
3. HIGH findings (should fix before merge)
4. MEDIUM findings (fix soon, can merge)
5. LOW findings (nice to have)
6. POSITIVE observations (what is done well — at most 3)
7. VERDICT: PASS / NEEDS_CHANGES / BLOCK
```
Rules:
- One vocabulary: the review's overall verdict uses only PASS / NEEDS_CHANGES / BLOCK (the vocabulary `scripts/validate-review-output.py` accepts). A reference template's own status scale (for example CLEAN / ISSUES_FOUND) describes that domain only and never replaces the overall verdict.
- CRITICAL findings automatically produce a BLOCK verdict.
- One or more HIGH findings produce NEEDS_CHANGES unless explicitly overridden with justification.
- A PASS verdict with zero findings requires a justification paragraph explaining what was checked and why nothing was found.
- Do not pad the POSITIVE section to soften a negative verdict. If nothing stands out positively, say "None noted."
## Tool Use
These rules are yours to follow. The worker runs as `general-purpose` with every tool, so nothing blocks a write in review mode.
### Review Mode (Default)
**CAN Use**: Read, Grep, Glob, Bash (read-only commands, git diff)
**CANNOT Use**: Edit, Write, Bash (state-changing commands)
### Fix Mode (--fix)
**CAN Use**: Read, Grep, Glob, Edit, Bash (including git commands and test runners)
**CANNOT Use**: Write (for new files, except test-analyzer which can create test files)
## Reference Loading Table
| Signal | Load These Files | Why |
|---|---|---|
| Convention compliance, style, CLAUDE.md | `reviewer-code/code-quality.md` | "code quality", "style review", "convention check" |
| Simplify code for clarity | `reviewer-code/simplifier.md` | "simplify", "reduce complexity", "readability" |
| Language-specific idioms (Go/Python/TS) | `reviewer-code/language-specialist.md` | "language idioms", "modern stdlib", "Go/Python patterns" |
| Naming conventions, casing drift | `reviewer-code/naming.md` | "naming consistency", "acronym casing", "convention drift" |
| Unreachable branches, unused exports | `reviewer-code/dead-code.md` | "dead code", "unused", "orphaned files" |
| Comment accuracy, staleness, quality | `reviewer-code/comments.md` | "comment accuracy", "comment rot", "stale comments" |
| Hot paths, N+1, allocations | `reviewer-code/performance.md` | "performance", "hot paths", "N+1", "allocations" |
| Type invariants, encapsulation | `reviewer-code/type-design.md` | "type design", "type safety", "illegal states" |
| Test coverage quality, gaps | `reviewer-code/test-analyzer.md` | "test coverage", "test quality", "test gaps" |
| Language-specialist review: full Go/Python/TypeScript check catalog | `reviewer-code/language-checks.md` | load with `language-specialist.md` |
| Hardcoded values, env vars, secrets | `reviewer-code/config-safety.md` | "config safety", "hardcoded values", "secrets in code" |
Related skills
| Skill | Stars | Updated | Works with | Status | Install |
|---|---|---|---|---|---|
| rea-analysis-optimizationmorluto/reaOptimize REA providers and shared analysis workflows for agent context, latency, CPU, and memory while preserving correctness and Evidence. Use for analysis bottlenecks and cache or query design; CI/test-suite optimization belongs to rea-ci-tests-audit. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| rea-ci-tests-auditmorluto/reaAudit REA CI and test suites for correctness, fixture drift, false passes, and measured performance; use for suite or workflow investigations and efficiency changes, rather than ordinary feature testing or a single test failure. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics. | Stars: 471k | Updated | Claude CodeCodex | Verified | Install |
| ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first. | Stars: 194k | Updated | Claude CodeCodex | Verified | Install |