# ReqForge

> From requirements to shippable products (primary). Open-source Agent Harness for Claude Code, Cursor, OpenCode, and Gemini CLI.

## One-Line Definition

**ReqForge** leads with the outcome: requirements → Spec → Plan → Build → Review → Release.

Under the hood it is an **Agent Harness** — constraints, guidance, and feedback (skills, hooks, memory, evolution) so AI output stays verifiable. Harness = Guides (feedforward) + Sensors (feedback) + Steering Loop (evolution).

## Problem It Solves

Vibe Coding without a harness: AI writes code → direction drifts → no design specs → bugs multiply → context grows → AI forgets earlier requirements → code diverges. The root cause isn't model capability — it's the lack of a **system** around the model.

## How It Works

### Workflow (0-to-1)

0. **Domain research (optional)** → `/domain-mapper` → domain-map.md (orthogonal to spec→build)
1. **Requirements** → `/product-spec-builder` → Product-Spec.md (Multi-Stakeholder Review + Critique Gate default; optional PM frameworks + CoT)
2. **Design Brief (optional)** → `/design-brief-builder` → Design-Brief.md
3. **Design mockups (optional)** → `/design-maker` → mockups + ephemeral UI-Spec.md
4. **Development Plan** → `/dev-planner` → DEV-PLAN.md
5. **Implementation** → `/dev-builder` → project code (one Phase per invocation)
6. **Review & Release** → `/code-review`, `/bug-fixer` (+ `pnpm forge-bug-fix`), `/release-builder`

### Routing & brownfield

- `/request-dispatcher` — when static Skill dispatch cannot pick one target
- `/change-manager` — one feature per `changes/<name>/`: propose → apply → verify → archive

### Core Architecture

- **14 Skills** — product-spec-builder, change-manager, design-brief-builder, design-maker, dev-planner, dev-builder, bug-fixer, code-review, release-builder, domain-mapper, request-dispatcher, feedback-writer, evolution-engine, skill-builder
- **10 Sub-Agents** — implementer, planner, code-reviewer + 4 specialists, feedback-observer, evolution-runner, test-writer
- **10 default hooks** — hallucination-gate, pre-commit-check, phase-exit-guard, stop-gate, retry-gate, detect-feedback-signal, mark-review-needed, check-evolution, memory-check, memory-guard
- **3-Tier Memory** — project-memory.md, decisions-log.md, task-history.md
- **Evolution Loop** — feedback → pattern detection (3+) → proposals → skill-builder
- **Loadouts** — `full`, `web-app`, `cli-tool`, `minimal` — see `core/docs/loadout-scenarios.md`

### Supported Clients

| Client | Adapter | Control file |
|--------|---------|--------------|
| Claude Code | `adapters/claude-code/` | `.claude/CLAUDE.md` |
| Cursor | `adapters/cursor/` | `.cursor/rules/` |
| OpenCode | `adapters/opencode/` | `.opencode/AGENTS.md` |
| Gemini CLI | `adapters/gemini-cli/` | `.gemini/GEMINI.md` |

Install: `pnpm forge-install <client> --target <dir>` · Sync core: `pnpm sync`

Docs: README.md · README.zh-CN.md · core/docs/ · Product-Spec.md

Version: v1.44.0
