vibe2prod 1.3.2

AI-powered code audit and remediation. Turns vibe-coded MVPs into production-ready software.

FORGE

Framework for Orchestrated Remediation & Governance Engine

A 12-agent AI system that scans codebases for security, quality, and architecture issues — then fixes them.

Quick Start

# Install
pip install vibe2prod

# Register as MCP server in Claude Code
claude mcp add forge -e OPENROUTER_API_KEY=your-key -- python -m forge.mcp_server

# Scan a repo
# (use the forge_scan tool in Claude Code)

Get an OpenRouter API key at openrouter.ai (free signup).

The /forge Skill

After scanning, use the /forge skill in Claude Code to autonomously fix findings. It reads the scan report, prioritizes issues, and applies fixes with micro-commits.

Full CLI documentation: vibe2prod.net/cli

Architecture

Discovery (Agents 1-4)     Scan codebase, identify issues
    |
Triage (Agents 5-6)        Classify by complexity tier (0-3), plan fixes
    |
Remediation (Agents 7-10)  Apply fixes via three control loops
    |
Validation (Agents 11-12)  Verify fixes, generate readiness report

Agents

#	Agent	Role
1	Codebase Analyst	Map architecture, files, dependencies
2	Security Auditor	3-pass parallel security scan
3	Quality Auditor	3-pass parallel quality scan
4	Architecture Reviewer	Structural coherence evaluation
5	Fix Strategist	Prioritize and order fixes
6	Triage Classifier	Assign complexity tiers (0-3)
7	Coder Tier 2	Scoped fixes (1-3 files)
8	Coder Tier 3	Architectural fixes (5-15 files)
9	Test Generator	Write tests for fixes
10	Code Reviewer	Review fix quality
11	Integration Validator	Verify merged codebase
12	Debt Tracker	Generate readiness report

Control Loops

• Inner Loop: Coder -> Review -> Retry (max 3 iterations)
• Middle Loop: Escalation when inner loop exhausted (RECLASSIFY / DEFER)
• Outer Loop: Re-plan with Fix Strategist (max 1 replan)

Tier Routing

• Tier 0: Auto-skip (invalid / false-positive)
• Tier 1: Deterministic fix (no LLM needed)
• Tier 2: Scoped AI fix (1-3 files, Sonnet 4.6)
• Tier 3: Architectural AI fix (5-15 files, Sonnet 4.6)

Requirements

• Python 3.12+
• OpenRouter API key (for LLM providers)
• AgentField control plane (optional — only needed for platform mode)

Usage

Standalone Mode

Run FORGE locally without an AgentField server:

from forge.standalone import run_standalone

result = await run_standalone(repo_path="./my-app", config={"mode": "discovery"})

AgentField Mode

# Start as AgentField node
python -m forge

# Or via entry point
forge-engine

FORGE registers as an AgentField node (forge-engine) and exposes three reasoners:

• remediate — Full pipeline: scan -> triage -> fix -> validate
• discover — Scan-only mode (Agents 1-6, no fixes)
• scan — Alias for discover (free tier)

Hive Discovery (Swarm Mode)

An alternative discovery architecture using a three-layer swarm approach:

config = {"discovery_mode": "swarm"}  # default: "classic"

See doc/hive-discovery-spec.md for the full design.

Configuration

Model routing is configurable per-agent via the models dict:

config = {
    "models": {
        "default": "anthropic/claude-haiku-4.5",
        "coder_tier2": "anthropic/claude-sonnet-4.6",
        "coder_tier3": "anthropic/claude-sonnet-4.6",
    }
}

Resolution: defaults < models.default < models.<role>

Resilience

FORGE normalizes LLM outputs before validation to handle model inconsistencies:

• Category aliases: LLM-returned categories are mapped to canonical categories (quality, reliability, security) via _CATEGORY_ALIASES
• Priority floor: Priorities < 1 are clamped to 1 before validation
• Dependency coercion: depends_on_finding_id returned as a list is coerced to a string