convez.ai
AgenteDevOps & InfraestruturaAvançado
antlss

GitLab Review Agent

antlss/gitlab-review-agent

Revisor de código IA com contexto real, multi-LLM e aprendizado contínuo

Ver no GitHub
BomPluginScore · de 0 a 100

Agent autônomo de revisão de código para GitLab que combina múltiplos LLMs (OpenAI, Anthropic, Gemini) com exploração real do repositório. Não lê apenas diffs genéricos: clona o repo, navega a base de código, entende convenções históricas e aprende com feedback para melhorar continuamente.

Instalação

$npx convez-ai antlss/gitlab-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
15
Forks
3
Acompanhando
0
Saúde
57%
Último commit
há 7 meses
Versão
—
Issues abertas
3
Contribuidores
2

Casos de uso

  • Automação de revisão de code em Merge Requests do GitLab
  • Enforcement de padrões e convenções do repositório
  • Análise profunda de impacto de mudanças em código existente

Veredito

Ferramenta robusta e bem pensada que vai além de linters triviais. A arquitetura de agente com loop de feedback e roteamento multi-LLM é production-ready para times que usam GitLab e querem revisor IA realmente contextual. Código em Go é uma boa escolha para performance.

README

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

🤖 AI Review Agent

More than just a .patch reader. AI Review Agent is an autonomous, context-aware code review assistant designed natively for GitLab Merge Requests. It clones your repository, reads the codebase, checks historical conventions, and engages in technical debates—just like a Senior Developer.


🚀 Why AI Review Agent?

Most open-source AI reviewers simply pipe your git diff into an LLM and spit out generic advice. We do things differently:

  • 🧠 True Contextual Awareness: We don't just read the diff. The agent is equipped with tools (read_file, search_code, multi_diff) to explore the actual codebase. If you modify a function signature, it can search where else it's used before commenting. No more "hallucinated" bugs.
  • ♻️ Self-Improving Feedback Loop: The system learns your project's conventions. A background Cron job consolidates historical AI reviews and human feedback into a tailored "Repository Best Practices" rulebook.
  • 🔀 Robust Multi-LLM Routing: Avoid vendor lock-in. Natively supports OpenAI (GPT-4o), Anthropic (Claude 3.7), and Google (Gemini 2.0) with load-balancing and fallback mechanisms. Mix and match models based on cost, rate limits, or language proficiency.
  • 💻 Interactive Local CLI: Don't want to spam your team with AI comments? Run the agent locally via CLI. It performs a dry-run review, displays the findings in your terminal, and lets you interactively select exactly which comments to push to GitLab.

🔍 How It Works: The "Deep Dive" Review Flow

  1. Trigger & Initialization: A webhook catches a Merge Request event (create/update). The agent checks whether the MR carries the configured review trigger label (default: ai-review). If the label is present, a job is queued asynchronously and the specific repository config (frameworks, model overrides) is loaded. A manual CLI trigger skips the label check and always proceeds.
  2. Smart Git Synchronization: The agent acquires a lock and shallow-fetches the target branch. It calculates a smart Base SHA to only process incremental new commits if the MR was reviewed previously, preventing noisy duplicate comments.
  3. Risk Scoring & Parsing: Modifed files are scored for risk. Highly modified or complex files are pre-loaded directly into the LLM context. Massive PRs (>150 files) are safely truncated and sampled by risk to protect your context window.
  4. Context Gathering (The Secret Sauce): External data is fetched:
    • Repository Settings: Known frameworks/languages.
    • Discussion History: Previous unresolved AI comments (it can auto-resolve them if the developer fixed the code!).
    • Feedback Rules: Historical lessons learned specific to this repo.
  5. Agentic Code Analysis: The LLM runs in an agentic loop. Over multiple iterations, it navigates the codebase using tools (read_file, search_code). It verifies its assumptions against real code before drafting a comment.
  6. Publish & Auto-Resolution: Validated, structured comments are pushed as inline GitLab discussions. If a developer modified lines overlapping with a previous AI comment, the agent automatically recognizes the fix and resolves the old thread.
  7. Reply Loop: Developers can reply directly to the AI's thread in GitLab. A specialized Replier Agent wakes up, reads the thread history + surrounding code context, and continues the technical debate.

⚡ Quick Start

Prerequisites

  • Go 1.25.5+
  • A GitLab instance (or gitlab.com)
  • Access Tokens: GitLab (Personal/Project Access Token) and at least one LLM Provider (OpenAI, Anthropic, Google).

Docker Installation (Recommended)

The easiest way to get the agent running without managing Go environments is using Docker Compose.

git clone https://github.com/antlss/gitlab-review-agent.git
cd gitlab-review-agent

# Configure your environment
cp .env.example .env
# Edit .env to set your LLM tokens and GitLab credentials

# Start the server in the background
docker-compose up -d

Your webhook server will be live at port 8080. You can execute the interactive CLI directly inside the running container:

docker exec -it ai_review_agent ./cli review --project-id 123 --mr-id 45

Manual Installation (From Source)

git clone https://github.com/antlss/gitlab-review-agent.git
cd gitlab-review-agent

# Build the server and CLI binaries
go build -o server ./cmd/server
go build -o cli ./cmd/cli

# Configure your environment
cp .env.example .env

Edit .env to define your GITLAB_BASE_URL, GITLAB_TOKEN, STORE_DRIVER (file or sqlite are easiest to start), and your preferred LLM_DEFAULT_PROVIDER.

Running the Server (Webhook Mode)

Start the webhook handler and background worker pool:

./server

Point your GitLab Project Webhook to http://<your-server>:8080/webhook/gitlab.

Label-Based Review Triggering

By default the agent only reviews Merge Requests that carry the ai-review label. Add the label to an MR in GitLab and the next open/update webhook event will trigger a review automatically.

You can change the label name via the REVIEW_TRIGGER_LABEL environment variable:

# .env
REVIEW_TRIGGER_LABEL=ai-review   # default — change to any GitLab label you prefer

Note: The CLI (./cli review) bypasses the label check entirely — it always performs a review as long as the configured GitLab token has access to the target project.

Running the CLI (Interactive Dry-Run Mode)

Trigger a review manually from your terminal and pick which comments to actually post:

./cli review --project-id 123 --mr-id 45

You can dynamically override the model for a specific run:

./cli review --project-id 123 --mr-id 45 --model claude-3-7-sonnet-20250219

🏗️ Architecture Overview

  • cmd/server: HTTP server handling GitLab Webhooks, Cron jobs, and Worker pools.
  • cmd/cli: The Command Line Interface for interactive local reviews.
  • internal/core: Heart of the logic (review and reply pipelines, feedback loops, reviewer / replier agents).
  • internal/pkg: External port adapters (GitLab API, Git CLI wrapper, LLM drivers, SQL/File storage DAOs).

We strictly follow Standard Go Project Layout conventions and utilize Clean Architecture principles.


🤝 Contributing

We welcome contributions! Please refer to the CONTRIBUTING.md for local development setup, coding standards, and our branch/PR workflow.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

Ver o original no GitHub

Bom para

Relacionados