convez.ai
AgenteAutomaçãoAvançado
Juadsuarezsan

Code Review Agent

Juadsuarezsan/code-review-agent

Revisor automático de PRs com detecção de bugs e análise de cobertura de testes

Ver no GitHub
InicialPluginScore · de 0 a 100

Agent de revisão de código que analisa pull requests usando tree-sitter, análise estática e Claude para detectar bugs, sugerir refatorações, identificar gaps de testes e gerar comentários inline automáticos. Integra-se com GitHub Actions e é avaliado contra SWE-bench Lite.

Instalação

$npx convez-ai Juadsuarezsan/code-review-agent

1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.

Métricas

Estrelas
0
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 1 mês
Versão
—
Issues abertas
0
Contribuidores
1

Casos de uso

  • Automatizar revisão de código em pull requests no GitHub
  • Detectar bugs e problemas de segurança antes do merge
  • Identificar gaps de testes e sugerir cobertura
  • Gerar comentários inline contextualizados com justificativas

Veredito

Projeto ambicioso que combina análise estática (tree-sitter, ruff, semgrep) com Claude para revisão semântica real. Ainda em fase de planejamento, mas a arquitetura é sólida e o roadmap claro — útil para times que querem CI/CD mais inteligente e menos ruído que ferramentas convencionais.

README

A tradução para o português deste README ainda está na fila.

Project 06 — Code Review Agent

Analyzes pull requests, detects real bugs, suggests refactorizations, writes missing tests, comments inline on the diff. The space behind Copilot Workspace, Cursor, Codium, Sourcegraph Cody.

Status LLM Bench

Industrial use case: Developer tools — Cursor, Codium, Sourcegraph Cody, Codacy.

What this project does

Reads a PR (diff + context), runs multiple analyzers in parallel (bug detector, style checker, test gap analyzer, security scanner, performance reviewer), synthesizes the top-N most actionable comments by severity. Posts inline comments to GitHub via Actions.

Architecture

PR (GitHub webhook or manual input)
   │
   ▼
[PR Analyzer] PyGithub → changed files, lines, context
   │
   ▼
[Context Builder]
   ├─ AST via tree-sitter
   ├─ related files (imports, importers)
   ├─ git history
   └─ linked issue
   │
   ▼
[Multi-aspect Analyzers] (parallel)
   ├─ Bug Detector (Claude + static analysis)
   ├─ Style Checker (ruff/eslint + Claude)
   ├─ Test Gap Analyzer (Claude)
   ├─ Security Scanner (semgrep + Claude)
   └─ Performance Reviewer (Claude)
   │
   ▼
[Comment Synthesizer] Claude → inline comments
   │ location, severity, suggestion, justification
   │
   ▼
[Priority Filter] → top N by severity
   │
   ▼
[Output] post to GitHub or return JSON

Roadmap to v1.0.0

  1. PyGithub integration for PR reading + commenting
  2. tree-sitter for multi-language AST (Python, JS/TS, Go)
  3. Static analysis wrappers (ruff, mypy, eslint, semgrep)
  4. LangGraph parallel analyzer pipeline
  5. Comment synthesizer with severity scoring
  6. Eval on SWE-bench Lite (300 issues) — bug detection F1
  7. LLM-as-judge on 100 generated comments (actionability, specificity, correctness)
  8. Next.js demo with diff viewer (react-diff-viewer)
  9. Gallery of 10 reviews on famous public PRs
  10. GitHub Action published to marketplace for self-hosted use

Stack

LayerTechnology
LLMClaude Sonnet 4.5
ASTtree-sitter (Python, JS/TS, Go bindings)
Static analysispylint, ruff, mypy (Python), ESLint (JS/TS)
Securitysemgrep
Diff parsingunidiff or difflib
GitHubPyGithub
OrchestrationLangGraph
StoragePostgreSQL for historical reviews
FrontendNext.js with react-diff-viewer
ObservabilityLangSmith

Definition of Done — project-specific

  • Pipeline runs over all 300 SWE-bench Lite issues
  • Bug detection F1 reported against ground truth fixes
  • Comment quality LLM-judged over 100 cases
  • Demo: paste a diff or PR URL, see review generated
  • Gallery of 10 reviews on famous public PRs in the demo
  • GitHub Action published to the marketplace for use in real repos
  • Cost-vs-quality analysis: when is this worth it vs ruff alone

Plus the 12 universal DoD blocks.

License

MIT.

Ver o original no GitHub

Bom para

Relacionados