Vibe

SPEC-driven AI Coding Framework with Multi-LLM Orchestration

AI coding framework for Claude Code. SPEC-based requirements management, Multi-LLM (Claude + GPT + Gemini) orchestration, and automated quality assurance with 13+ parallel review agents.

Quick Start

npm install -g @su-record/vibe
vibe init

Core Workflow

/vibe.spec → /vibe.spec.review → [SPEC Summary] → /vibe.run → /vibe.trace → /vibe.review → [Retrospective]
     ↓              ↓                  ↓               ↓            ↓              ↓              ↓
  Write SPEC   GPT/Gemini Review  User Review     Execute     Traceability   Parallel Review  Auto-Save

Workflow Details

| Step | Command | What Happens | |------|---------|-------------| | 1 | /vibe.spec "feature" | Generate SPEC + parallel research agents | | 2 | /vibe.spec.review "feature" | 3-round GPT/Gemini cross-validation | | 3 | SPEC Summary | User reviews spec before implementation | | 4 | /vibe.run "feature" | Execute implementation phase-by-phase | | 5 | /vibe.trace "feature" | Requirements traceability matrix | | 6 | /vibe.review | 13+ parallel review agents | | 7 | Auto-Retrospective | Lessons learned saved for cross-session recall |

Key Features

| Feature | Description | |---------|-------------| | SPEC-driven Development | Traceable: Requirements → SPEC → Feature → Test | | Multi-LLM Orchestration | Claude + GPT + Gemini 3-way verification | | 13+ Parallel Review Agents | Security, Performance, Architecture, etc. | | ULTRAWORK Mode | All optimizations with a single keyword | | Smart Hook Dispatcher | Pattern-matched prompt routing (no wasted LLM calls) | | Progress Tracking | claude-progress.txt survives context compaction | | Auto-Retrospective | Lessons learned auto-saved after /vibe.run | | Swarm Pattern | Auto-split complex tasks for parallel processing | | 23 Language Presets | TypeScript, Python, Go, Rust, Swift, Kotlin, and more |

Commands

Terminal

| Command | Description | |---------|-------------| | vibe init | Initialize project | | vibe update | Update configuration | | vibe status | Check status | | vibe gpt auth | GPT OAuth authentication | | vibe gemini auth | Gemini OAuth authentication |

Claude Code Slash Commands

| Command | Description | |---------|-------------| | /vibe.spec "feature" | Generate SPEC document + parallel research | | /vibe.run "feature" | Execute implementation | | /vibe.run "feature" ultrawork | Maximum performance mode | | /vibe.verify "feature" | BDD verification | | /vibe.trace "feature" | Requirements traceability matrix | | /vibe.review | 13+ agent parallel review | | /vibe.review --race | GPT + Gemini race review |

ULTRAWORK Mode

Activate maximum performance with ultrawork or ulw keyword:

/vibe.run "feature" ultrawork

Enabled Features:

• Parallel sub-agent exploration (3+ concurrent)
• Background agents (prepare next phase during implementation)
• Phase pipelining (eliminate wait time between phases)
• Boulder Loop (auto-continue until all scenarios complete)
• Auto-save at 80%+ context

Performance:

| Mode | Relative Speed | |------|---------------| | Sequential | Baseline | | Parallel | ~25% faster | | ULTRAWORK + Pipeline | ~50% faster |

Multi-LLM Orchestration

Smart Hook Dispatcher

A single dispatcher reads user prompts and routes to the right LLM only when patterns match. No wasted API calls:

| Pattern | Routes to | Use Case | |---------|-----------|----------| | Architecture keywords | GPT | Architecture review | | UI/UX keywords | Gemini | UI/UX feedback | | Debug keywords | GPT | Bug analysis | | Code analysis keywords | Gemini | Code quality review | | No match | None | No external LLM call |

Race Review

Run GPT + Gemini in parallel for cross-validation:

/vibe.review --race

| Agreement | Priority | Action | |-----------|----------|--------| | Both agree (100%) | P1 | High confidence | | One model (50%) | P2 | Needs verification |

Parallel Review Agents

Run 13+ agents simultaneously with /vibe.review:

| Category | Agents | |----------|--------| | Security | security-reviewer, data-integrity-reviewer | | Performance | performance-reviewer, complexity-reviewer | | Architecture | architecture-reviewer, simplicity-reviewer | | Language | python, typescript, rails, react reviewers |

Priority System:

• P1 (Critical): Blocks merge
• P2 (Important): Fix recommended
• P3 (Nice-to-have): Backlog

Swarm Pattern

Automatically split complex tasks for parallel processing:

Prompt → Complexity Analysis → Split Decision
                ↓
    Low  → Direct execution
    High → Create subtasks → Parallel processing → Merge results

Tasks are scored by complexity. When the score exceeds the threshold (default: 15), the task is automatically split into subtasks and processed in parallel.

Requirements Traceability

Track from requirements to tests with /vibe.trace:

REQ-login-001 → SPEC Phase 1 → Feature Scenario 1 → login.test.ts
REQ-login-002 → SPEC Phase 2 → Feature Scenario 3 → auth.test.ts

Project Structure

Global (~/.claude/):

~/.claude/
├── commands/     # Slash commands
├── agents/       # Review/research agents
├── skills/       # Auto-activated guides
└── settings.json # Hooks

Project (.claude/vibe/):

.claude/vibe/
├── specs/              # SPEC documents
├── features/           # BDD scenarios
├── retros/             # Auto-retrospectives
├── progress.json       # Structured progress state
├── claude-progress.txt # Human-readable progress (for context survival)
├── config.json         # Project settings
└── constitution.md

Code Quality Standards

| Metric | Limit | |--------|-------| | Function length | 30 lines (recommended), 50 lines (allowed) | | Nesting depth | 3 levels | | Parameters | 5 | | Cyclomatic complexity | 10 |

API Usage

// Background agent
import { launch, poll } from '@su-record/vibe/orchestrator';
const { taskId } = launch({ prompt: 'Analyze code', agentName: 'analyzer' });
const result = await poll(taskId);

// Swarm pattern - Auto-split complex tasks
import { swarm } from '@su-record/vibe/orchestrator';
const result = await swarm({
  prompt: 'Implement login with: 1. UI 2. Validation 3. API 4. Tests',
  maxDepth: 2,
  splitThreshold: 15,
});

// LLM direct call
import { ask } from '@su-record/vibe/lib/gpt';
import { webSearch } from '@su-record/vibe/lib/gemini';

IDE Support

Cursor

Assets auto-installed on vibe init/update:

| Asset | Path | Count | |-------|------|-------| | Subagents | ~/.cursor/agents/ | 12 | | Skills | ~/.cursor/skills/ | 7 | | Rules | ~/.cursor/rules-template/ | 23 languages |

Requirements

• Node.js 18.0.0+
• Claude Code

License

MIT - GitHub