convez.ai
AgenteAutomaçãoAvançado
marcoantonios1

Forge

marcoantonios1/Forge

Agent IA para coding no terminal com suporte a múltiplos LLMs

Ver no GitHub
InicialPluginScore · de 0 a 100

Agente autônomo de engenharia de software para terminal, alimentado por múltiplos LLMs via Costguard. Lê seu repositório, compreende tarefas em linguagem natural, gera patches e aguarda confirmação antes de aplicar mudanças. Model-agnostic: funciona com modelos locais (Ollama) ou na nuvem (Anthropic, OpenAI).

Instalação

$npx convez-ai marcoantonios1/Forge

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
1
Forks
0
Acompanhando
0
Saúde
42%
Último commit
há 4 meses
Versão
—
Issues abertas
0
Contribuidores
2

Casos de uso

  • Automatizar tarefas de desenvolvimento repetitivas descrevendo-as em linguagem natural
  • Gerar patches validados com preview antes de aplicar ao repositório
  • Integrar diferentes provedores de LLM mantendo controle de custos via Costguard

Veredito

Ferramenta sólida e bem arquitetada para automação de tarefas de código via CLI, com fluxo de confirmação seguro e suporte a múltiplos backends de IA. Requer setup de Costguard e Go 1.22+, sendo mais adequada para equipes técnicas que buscam agência local controlada.

README

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

Forge

An autonomous software engineering agent for the terminal. You describe a task in plain English; Forge classifies it, reads your repository, writes a patch, and asks for confirmation before touching anything.

> add input validation to the login handler
compiled: feature/module [supervised]
  ⚙  Starting task  feature/module  [a3f2c1b0]
  →  git_status  (repo)
  ✓  git_status  on branch main, clean
  →  read_file  internal/auth/handler.go
  ✓  read_file  read 84 lines
  →  search_code  validateInput
  ✓  search_code  found 0 matches
┌─ Patch preview ────────────────────────────────┐
│  1 file(s) will be modified                      │
└──────────────────────────────────────────────────┘
Apply this patch? [y]es / [n]o / [a]ll session  y
  ✔  Applied  internal/auth/handler.go  (1 hunks)
╔══════════════════════════════════════════╗
║  Task complete                           ║
╚══════════════════════════════════════════╝
   Added input validation to the login handler.

How it works

Your input
    ↓
Task Compiler      — classifies intent, scope, and execution policy
    ↓
Agent Loop         — reads the repo with read_file, list_files, search_code, git_*
    ↓
Patch System       — generates a unified diff, validates it, and applies atomically
    ↓
Confirmer          — shows a diff preview and prompts before any file is written

All model calls are routed through Costguard, an OpenAI-compatible proxy that handles routing, budget enforcement, and cost tracking. No provider is contacted directly.

Requirements

  • Go 1.22+
  • A running Costguard instance
  • A model served via Costguard (Ollama, Anthropic, OpenAI, etc.)

Installation

git clone https://github.com/marcoantonios1/Forge
cd Forge
make build
# binary is at ./bin/forge

Configuration

Copy .env.example to .env and fill in your values:

cp .env.example .env
VariableDefaultDescription
COSTGUARD_URLhttp://localhost:8080Costguard proxy URL
COSTGUARD_MODEbalancedRouting mode: cheap, balanced, best, private
COSTGUARD_AGENTforgeAgent identity header sent to Costguard
COSTGUARD_PROVIDER(unset)Backend provider: local_ollama, anthropic, openai, …
COSTGUARD_TEAM(unset)Team label for cost attribution
COSTGUARD_PROJECT(unset)Project label for cost attribution
COSTGUARD_TIMEOUT60sPer-request timeout
COSTGUARD_MAX_RETRIES3Retries on 429 / 502 / 503
COMPILER_MODELclaude-haiku-4-5-20251001Model used for task classification and agent loop
FORGE_DEBUGfalseVerbose logging

Environment variables take precedence over .env.

Usage

./bin/forge [--debug] [--yes]
FlagDescription
--debugEmit structured JSON events instead of formatted output
--yesApprove all patches without prompting (autonomous mode)

forge.md

Drop a forge.md in your project root to give Forge persistent project context — build commands, style rules, frozen modules, banned patterns. It is read at startup and prepended to every agent system prompt.

# forge.md
build: make build
test: make test
style: always use tabs, never spaces
banned: fmt.Println in production code
notes: internal/auth is frozen — do not modify

Execution policies

The task compiler assigns one of three policies based on the task description:

PolicyBehaviour
autonomousAgent runs and applies patches without confirmation
supervisedAgent runs; you confirm each patch
safeAgent runs; confirmation required; used for deletions, secrets, infra changes

--yes overrides all policies and bypasses confirmation.

Architecture

cmd/forge/
    main.go               entry point, REPL, flag parsing

internal/
    config/               env + .env loader
    session/              crypto/rand session IDs
    projectconfig/        forge.md loader
    compiler/             natural language → typed Task struct
    costguard/            OpenAI-compatible HTTP client with retry/backoff
    agent/                control loop, tool registry, system prompt
    tools/                read_file, list_files, search_code, git_status, git_diff, git_log
    patch/                unified diff parser, validator, atomic applier, undo history
    confirm/              AutoConfirmer, NullConfirmer, SafeConfirmer
    events/               event types and Emitter interface
    ui/                   terminal renderer, diff coloriser, TTY detection

Development

make build    # compile to ./bin/forge
make run      # build and run
make clean    # remove ./bin

License

MIT

Ver o original no GitHub

Bom para

Relacionados