Skip to content

notque/vexjoy-agent / skills/review/review

review

Verified

Code review: systematic single-file, parallel multi-reviewer, full-repo audit, PR diff review.

Review

Three modes. Pick one by request shape, default to systematic.

Request matchesMode
Review a file, diff, or PRSystematic (default)
"parallel review", "multi-reviewer", "3-reviewer"Parallel
"full repo review", "codebase audit", "repo health", review ALL filesFull-repo

Deep References

SignalLoadWhy
Reviewing Go codereferences/go-review-patterns.mdExports, concurrency, resources, metrics, tests
Dispatching Architecture reviewer in parallel modereferences/architecture-smell-baseline.md12 Fowler smells with language counter-examples and severity cap
Writing full-repo reportreferences/report-template.mdReport structure and field definitions
Dispatching full-repo wave agentsreferences/audit-playbook.md8-category checklists with evidence requirements
Disproving findings before report (Systematic Phase 3.5, Parallel Step 3)references/disprove-pass.mdRefuter dispatch, drop rule, fail-open cases, model choice
Receiving review feedbackreferences/receiving-feedback.mdFeedback-handling patterns

Severity Classification

Shared across all modes. When in doubt, classify UP.

LevelScopeExamples
BLOCKINGSecurity, correctness, reliabilityAuth bypass, race condition, resource leak, logic error, test failure
SHOULD FIXMaterial quality, patterns, testsMissing tests, unhelpful errors, pattern violations, N+1 in hot paths
SUGGESTIONOptional, stylisticNaming preferences, comments, micro-optimizations

Decision: security/correctness/reliability risk? -> BLOCKING. Violates patterns or creates maintenance burden? -> SHOULD FIX. Purely stylistic? -> SUGGESTION.

Verdict Rules

ConditionVerdict
Any BLOCKING (parallel: CRITICAL) findingBLOCK
SHOULD FIX (parallel: HIGH) findings, no BLOCKINGNEEDS_CHANGES
Only SUGGESTION or cleanPASS

These are the only verdict words, the same PASS / NEEDS_CHANGES / BLOCK the reviewer agents emit. An open question that stops a PASS is NEEDS_CHANGES with the question stated. The validator rejects a verdict that contradicts the findings (PASS over SHOULD FIX, anything but BLOCK over BLOCKING).

Omit empty severity sections. Include review scope, evidence, and limitations. Validate output (--type parallel for Mode 2):

python3 ~/.claude/scripts/validate-review-output.py --type systematic /tmp/review-output.md

Exit 0 = valid; 1 = schema errors; 2 = unparseable; 3 = missing jsonschema.


Mode 1: Systematic Review (default)

Single-reviewer, 4-phase review of a named file, diff, or PR.

Phase 1: UNDERSTAND

Read the repo's diff-triggered checklist from the default branch (git show origin/main:REVIEW.md) and apply matching rows to the diff. Read repository instructions and the complete diff. Read surrounding code to understand each changed path and its consumers. When signatures change, find all callers and interface implementations. Trace parameters to their source: query params may hold any user string; tokens may be server-issued IDs; enums have bounded values. Check validation at each caller.

Gate: Every changed path accounted for, callers traced where needed.

Phase 2: VERIFY

Run relevant tests and required repository checks. Reuse prior results only when they cover current code and environment. Check material claims in comments and PR description against code and tests. Missing tools, skipped checks, and inferred outcomes are not passes.

Gate: Claims have supporting evidence; required checks passed or the gap is recorded.

Phase 3: ASSESS

Assess risks: security (auth, validation, injection, secrets), performance (N+1, unbounded work, allocations on hot paths), architecture (conventions, compatibility, scope, unnecessary abstractions). For extracted helpers, recheck contract and callers.

Gate: Relevant risks and remaining uncertainty are explicit.

Phase 3.5: VERIFY FINDINGS

Before reporting, check each finding's input sequence, existing guards, and proposed fix. Drop unreachable or already-handled claims. Deduplicate when combining reviews; resolve severity from evidence. Then run the disprove pass in references/disprove-pass.md: it drops findings refuted by quoted code and keeps the rest.

Phase 4: DOCUMENT

Review Summary:
  Files Reviewed: N | Lines Changed: +X/-Y
  Test Status: PASS | FAIL | SKIPPED
  Risk Level: LOW | MEDIUM | HIGH | CRITICAL

BLOCKING:
  1. Issue and consequence — path/file.ext:42
SHOULD FIX:
  1. Issue and consequence — path/file.ext:52
SUGGESTIONS:
  1. Optional improvement — path/file.ext:62

Verdict: PASS | NEEDS_CHANGES | BLOCK
Rationale: Evidence, scope, and limitations.

Mode 2: Parallel Review

Three concurrent reviewers, then aggregate.

Step 1: Scope

Identify the diff or files. Record the reviewed revision.

git diff --name-only HEAD
gh pr view --json files -q '.files[].path'

Select Architecture reviewer by language: Go -> golang-general-engineer; Python -> python-general-engineer; TypeScript -> typescript-frontend-engineer; mixed -> Explore.

Step 2: Dispatch

Dispatch three reviewers together. Read-only; no code edits.

ReviewerFocus
SecurityAuth, authorization, input validation, secrets, OWASP
Business LogicRequirements, edge cases, state transitions, failure modes
ArchitectureDesign, structure, performance, maintainability, scope

Pass references/architecture-smell-baseline.md verbatim to the Architecture reviewer. Require [Reviewer] file:line format with severity and consequence.

Step 3: Aggregate and Verdict

Deduplicate findings, resolve severity disagreements from evidence. Run the disprove pass in references/disprove-pass.md on the merged list. Apply shared verdict rules. Output a single combined report with severity matrix, combined findings, summary by reviewer, and a recommendation line **BLOCK**, **NEEDS_CHANGES** or **PASS** followed by - and the rationale.

Gate: Every reviewer returned, or missing coverage is explicitly stated without claiming approval.


Mode 3: Full-Repo Review

All source files through a 4-wave comprehensive review. Produces a prioritized backlog, not auto-fixes. Use for quarterly health checks, post-refactor audits, or new codebase onboarding.

Options: --directory [dir] (scope to one dir), --skip-precheck, --min-severity [level].

Step 1: Discover and Pre-check

Scan all source files (scripts/, hooks/, skills/, agents/, docs/). Never fall back to git diff.

python3 ~/.claude/scripts/score-component.py --all-agents --all-skills --json

Save scores as triage context. A score alone never determines severity.

Step 2: Run Comprehensive Review

Call comprehensive-review with --review-only and the full file list. Run all 4 waves (0-3). Load references/audit-playbook.md as prompt context for wave agents.

Step 3: Report

Merge deterministic scores with LLM findings. Identify systemic patterns (3+ files). Write full-repo-review-report.md using references/report-template.md.

Gate: Report exists with severity sections and deterministic scores.


Error Handling

ErrorSolution
Reviewer times out or returns nothingReport partial findings, note gap, retry on reduced scope
Disprove pass fails or returns unparseable outputKeep all findings, note the gap in the verdict
Validator script missingRun review without validation, note gap in verdict
score-component.py failsProceed with LLM review only, note gap in report
Too many files for single sessionSplit by directory: scripts/, hooks/, agents/, skills/
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: 468kUpdated Claude CodeCodexVerifiedInstall
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: 468kUpdated Claude CodeCodexVerifiedInstall
rea-tool-designmorluto/reaAudit, investigate, design, or change REA investigation tools, CLI/MCP contracts, provider capabilities, and Evidence semantics.Stars: 468kUpdated Claude CodeCodexVerifiedInstall
ponytail-auditDietrichGebert/ponytailQuality audit of the whole repo: bugs, security, real load, missing tests, speed, and what to delete. Most important first.Stars: 194kUpdated Claude CodeCodexVerifiedInstall