convez.ai
MCP ServerAutomaçãoAvançado
srnichols

Plan Forge

srnichols/plan-forge

SDLC orchestration shop para AI coding agents com governance e memory

Ver no GitHub
BomPluginScore · de 0 a 100

Servidor MCP que orquestra o ciclo de vida completo de desenvolvimento AI-nativo, transformando ideias brutas em planos validados e código entregável através de um pipeline de 7 etapas com 16 agentes especializados. Integra planejamento, validação, execução hardened, monitoramento de drift em tempo real e aprendizado contínuo via memória compartilhada OpenBrain.

Instalação

$npx convez-ai srnichols/plan-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
6
Forks
2
Acompanhando
0
Saúde
100%
Último commit
há 2 semanas
Versão
v3.35.9
Issues abertas
27
Contribuidores
2

Casos de uso

  • Coordenar múltiplos agentes AI em projetos de código complexos com validação de gates
  • Implementar SDLC robusto sobre GitHub Copilot mantendo controle de custos e qualidade
  • Monitorar drift, secrets e dependências em código gerado por AI em produção
  • Construir memória institucional reutilizável de decisões de desenvolvimento

Veredito

Ferramenta madura e altamente sofisticada para quem precisa de SDLC enterprise-grade sobre AI agents. Exige conhecimento técnico profundo, mas resolve problemas reais de governança, reprodutibilidade e custo que GitHub Copilot não endereça. Recomendado para equipes que já dominam fluxos CI/CD e querem escalar AI coding com segurança institucional.

README

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

Plan Forge

The AI-Native SDLC Forge Shop

Plan Forge is the orchestration harness that sits on top of GitHub Copilot (and other AI coding tools). It does not replace your model or your IDE — it adds the SDLC layer GitHub deliberately leaves to the ecosystem: planning, validation gates, memory, cost control, and reviewer separation.

It is also licensed MIT because your SDLC is yours, and your institutional memory lives in OpenBrain — a user-owned service — because your accumulated decisions should not be trapped inside any one AI vendor.

A blacksmith doesn't hand raw iron to a customer. They smelt it, hammer it, temper it — and then they watch, because a blade that isn't maintained will dull.

Plan Forge is a full-lifecycle AI development shop. Raw ideas are smelted through the Crucible into structured plans. Plans are forged into working code through a 7-step hardened pipeline. Shipped code is guarded by LiveGuard — drift, secrets, dependencies, incidents, all watched in real time. And every finding is learned back into the shop's memory, so the next run starts smarter than the last.

Smelt the idea. Forge the code. Guard the build. Learn from every run.

License: MIT

Website · Shop Tour · Manual · Documentation · FAQ · Extensions · Spec Kit Interop

107 MCP Tools (primary pforge-mcp) + companion pforge-master tools · 48+ CLI Commands · 103+ REST Endpoints · ~21 Agents · 4 Reviewer Skills · 9 Presets · 7 Adapters · 7947 Tests · v3.20.0-dev

The Four Stations

Plan Forge is one shop with four stations. Each one handles a distinct part of the software lifecycle — and they all share one memory.

StationVerbWhat happens hereStart with
🪨 SmeltSmelt the ideaRough idea → Crucible interview → tempered plan with scope contract and validation gates.Crucible · Tempering design notes
🔨 ForgeForge the codeHardened plan → DAG-scheduled execution → quorum consensus → shipped code. The classic 7-step pipeline.Your First Plan · AI Plan Hardening Runbook
🛡️ GuardGuard the buildShipped code → drift scoring, secret scan, dep watch, regression guard, incident capture, remote alerts.What is LiveGuard? · LiveGuard Tools
🧠 LearnLearn from every runFindings → Hallmark provenance-stamped OpenBrain memory → capability-negotiated OpenBrain writes → Health DNA → self-tuned escalation, cost, and quorum thresholds. Anvil deduplicates high-frequency writes; Lattice indexes code structure for blast-radius scoring.Memory Architecture · Bug Registry + Testbed

And the control room that ties them together: the live dashboard at localhost:3100/dashboard with 25 tabs, session replay, WebSocket event hub, cost reports, OTLP traces, and a remote bridge for Telegram / Slack / Discord / OpenClaw.


Start Here

You are...Start with
Evaluating Plan ForgeRead the Shop Tour → Skim capabilities
A developer using VS Code + CopilotRun Quick Start → Read COPILOT-VSCODE-GUIDE.md
An AI agent setting up a projectRead AGENT-SETUP.md (your entry point)
Already shipping and want the watch layerJump to LiveGuard — runs standalone on any codebase
Just browsingKeep reading — or visit planforge.software

Verified Results

Verified: 38+ phases self-built, 7947 self-tests, 107 MCP tools on primary pforge-mcp (plus companion pforge-master tools), zero manual rollbacks. See docs/capabilities.md.

A/B Test Results (April 2026)

Same app, same model (Claude Opus 4.6), same time (~7 min). Only difference: Run A had Plan Forge.

MetricPlan ForgeVibe Coding
Tests6013
Interfaces60
DTOs90
Quality Score99/10044/100

Read the full results →


How the Stations Work Together

The four stations form a closed loop: Smelt intake → Forge builds → Guard watches → Learn captures findings into memory → next Smelt starts smarter. Inside the Forge, a deterministic slice executor runs the plan; a reflective inner loop — retries with reflexion context, trajectory capture, auto-skill promotion, and postmortems — turns every slice into a research step that teaches the next one. Together, the ten opt-in inner-loop subsystems compose into a self-deterministic agent loop that keeps execution reproducible while the loop's context improves each pass.

For architecture diagrams, the 7-step pipeline mermaid, the LiveGuard cycle, and the self-improving feedback loops, see the Shop Tour and the interactive manual.


The Virtual Engineering Team

Plan Forge is not "AI plus a code-completion plugin." It is a full enterprise engineering shop where every traditional role is filled by a specialized agent or guardrail, governed by 40 years of software engineering practice encoded into 17+ auto-loading instruction files and 20 specialized reviewers.

Traditional engineering team rolePlan Forge equivalent
Product Owner (spec, direction, acceptance)You — non-negotiable, non-replaceable
IC engineers (architecture, security, performance, DB, deploy, API, accessibility, multi-tenancy, CI/CD, observability, dependency, compliance)20 specialized agents + 17 auto-loading guardrail files
Tech lead / staff engineerQuorum mode (multi-model consensus) + auto-escalation on slice failure
Engineering managerForge-Master Observer + Auditor — agents supervising agents, not metaphorically, literally
QA teamTempering harness + testbed + regression guard + slice gates
SRE / on-callLiveGuard (secret scan, drift, dep watch, env diff, incident capture, runbooks)
Continuous improvement / retrosAudit loop + bug registry auto-smelt + Reflexion lessons + auto-skill promotion
Architecture review boardIndependent Session-3 review gate (fresh AI session, full guardrail load)
Institutional knowledge / wikiOpenBrain L3 memory with Hallmark provenance
Release managerShipper agent + release-checklist + version.instructions.md

Your three jobs as the human:

  1. Make a few technology choices once — pick a preset (dotnet, typescript, python, etc.), declare Project Principles, set forbidden patterns. One-time, then locked.
  2. Own the spec and the direction — Product Owner duties. Crucible interviews you; you decide what to build and why. The shop won't deviate from the Scope Contract you sign off on.
  3. Run manual acceptance testing — agents don't read minds. They implement what the spec says, not what you intended. A green pipeline (all tests pass, no drift, no secrets, no regressions, Forge-Master independently audited) proves the code is correct. Only you can decide whether it's what you actually wanted. That's Product Owner UAT, same as it has been for 40 years.

Three closed loops, not one. The Forge builds. Forge-Master supervises. The Learn station feeds every finding back into the next plan. The audit loop ran for two weeks unattended on a real production Next.js site and surfaced 30+ defects the maintainer didn't know existed — see Day in the Forge. Skeptical CTOs: this is the part that distinguishes Plan Forge from "an AI that writes code."


Quick Start

Prerequisites

  • VS Code with GitHub Copilot (free, Pro, or Enterprise)
  • Git installed
  • A shell for the setup script: on Windows, the PowerShell that ships with Windows (5.1) or PowerShell 7+, with no extra install; on macOS and Linux, Bash (the Bash 3.2 that ships with macOS is enough) or PowerShell 7+

1. Clone and Run Setup

git clone https://github.com/srnichols/plan-forge.git my-project-plans
cd my-project-plans
# Windows (PowerShell)
.\setup.ps1 -Preset dotnet          # or: typescript, python, java, go, swift, rust, php, azure-iac

# Mac / Linux
./setup.sh --preset dotnet

Setup copies all framework files, installs MCP dependencies, and generates config. Zero manual steps.

2. Start Planning

  1. Open VS Code → Copilot Chat → Agent Mode
  2. Describe your feature → the pipeline guides you through 7 steps
  3. LiveGuard watches automatically after you ship

3. (Recommended) Enable Persistent Memory

Plan Forge ships with L1 (Hub) + L2 (.forge/*.jsonl) memory built in. The L3 layer — cross-session, cross-tool, semantic-search memory that powers Reflexion lessons, Auto-skills, cross-project Federation, and 28 auto-capturing MCP tools — requires OpenBrain, a self-hosted MCP server (PostgreSQL + pgvector). Plan Forge works without it (every hook degrades silently), but the inner loop only improves over time when L3 is present.

Pick the path that fits:

PathTimeCostBest for
Docker Compose~5 minFreeLocal dev, single machine
Supabase Cloud~10 min~$0.10–$0.30 / moSolo / small team, zero ops
Kubernetes / Azure Container Apps~30 minCloud ratesTeams, federation across repos
Skip for now0 min—Try Plan Forge first; enable later with pforge brain hint

Full walkthrough: srnichols.github.io/OpenBrain. Already running OpenBrain? pforge brain status confirms Plan Forge sees it.

See docs/CLI-GUIDE.md for all presets, flags, and multi-agent options.


What's Included

9 Tech-Stack Presets

PresetStackPresetStack
dotnet.NET / C# / ASP.NET CoreswiftSwift / SwiftUI / Vapor
typescriptTypeScript / React / NoderustRust / Axum / Tokio
pythonPython / FastAPI / DjangophpPHP / Laravel / Symfony
javaJava / Spring Bootazure-iacBicep / Terraform / azd
goGo / Chi / Gin

7 AI Agent Adapters

One setup command, every tool: setup.ps1 -Agent all

GitHub Copilot (primary) · Claude Code · Cursor · Codex CLI · Gemini CLI · Windsurf · Generic

MCP Server (107 Tools)

pforge-mcp/server.mjs exposes core, LiveGuard, Watcher, Crucible, Tempering, Bug Registry, Testbed, Forge-Master, Hallmark, Anvil, Lattice, Sync, and Memory operations. Live dashboard at localhost:3100/dashboard. 103+ REST endpoints for external integrations.

Key tools: forge_run_plan · forge_liveguard_run · forge_analyze · forge_master_ask · forge_capabilities · forge_smith · forge_cost_report · forge_lattice_query · forge_sync_memories · forge_memory_capture

Optional Capabilities

FeatureHow to EnableWhat It Does
Quorum modeAutomatic (complexity ≥ 6)3 models analyze in parallel, reviewer synthesizes. Works OAuth-only via the Copilot CLI — no API keys required. Add XAI_API_KEY to mix in a Grok leg. Self-tuning threshold.
Audit Looppforge audit-loop or .forge.json#auditClosed-loop drain: content-audit scanner → triage → fix. Default off; opt-in via audit.mode: "auto" or "always".
Auto-escalationBuilt-inModel fails → auto-promotes. Chain reorders by success rate.
Cost trackingBuilt-inPer-slice tokens, 23-model pricing, --estimate with historical calibration.
OpenBrain memory (L3)Recommended — see Quick Start § 3The L3 memory layer. 28 tools auto-capture findings with Hallmark provenance stamps via capability-negotiated OpenBrain writes. Anvil deduplicates Δ-identical writes. 4 prompts search before acting. Without it, Reflexion, Auto-skills, and Federation are inert.
Extensionspforge ext add <name>HIPAA, SaaS multi-tenancy, etc.
CI validationsrnichols/plan-forge-validate@v1GitHub Action for plan quality gates.
NotificationsConfigure in .forge.jsonSlack, Discord, Telegram, webhooks via bridge.
Spec Kit bridgeAuto-detectedImport specs + constitution from Spec Kit projects.

Documentation

ResourcePurpose
docs/COPILOT-VSCODE-GUIDE.mdVS Code + Copilot walkthrough
docs/CLI-GUIDE.mdpforge CLI reference
docs/REST-API.mdAll 103 REST endpoints, organized by domain
docs/capabilities.mdFull feature reference — all 107 MCP tools, ~21 agents, 4 reviewer skills, 103+ REST endpoints
CUSTOMIZATION.mdAdapt guardrails for your project
planforge.software/manual/Interactive web manual (30 chapters + 14 appendices)
planforge.software/faq.htmlFAQ
AGENT-SETUP.mdAI agent entry point
CONTRIBUTING.md · docs/RELEASE-CHECKLIST.mdMaintainer guide · release & distribution-sync checklist

Git Workflow

git commit -m "<type>(<scope>): <description>"   # feat, fix, refactor, test, docs, chore

See .github/instructions/git-workflow.instructions.md for conventions.


Contributing

See CONTRIBUTING.md. For extensions: extensions/PUBLISHING.md. For skills: docs/SKILL-BLUEPRINT.md.


License

MIT — use these guardrails in your projects, teams, and tools.

Ver o original no GitHub

Bom para

Relacionados