Artibot

Cognitive orchestration OS for Claude Code with hierarchical memory, RLVR self-learning, MCP server, and multi-platform agent teams.

This repository ships two complementary plugins under one marketplace:

| Plugin | Target | Version | Best for | |---|---|---|---| | artibot | Claude Code (developer CLI) | 4.44.0 | full Agent Teams orchestration, TDD, code review, security audits, RLVR learning, MCP server, Goal-driven autopilot, /learning diagnostics, /save + /resume single-shot handoff, /go → /orchestrate build-sequence hand-off, ambient conversation ledger (no-command capture), safety boost (machineId frontmatter, git-lock fail, 10m throttle) | | artibot-cowork | Claude Cowork (knowledge workers) | 3.1.0 | marketing campaigns, long-form writing, AEO/GEO content, KR-market SEO, AI-slop detection, Claude Design, Routines, Ultraplan, Monitor |

Both plugins share the same DEV protocol, Korean market expertise, data-sovereignty policy, and 6-stage content quality pipeline. They differ only in target environment and skill mix.

A cognitive orchestration plugin for Claude Code powered by dual-process routing, lifelong learning, and the native Agent Teams API. Artibot uses a System 1/2 cognitive architecture to classify every request, assembles specialized agent teams with P2P communication, and continuously learns from session outcomes to improve routing accuracy over time.

<p align="center"> <img src="assets/artibot-mascot.jpeg" alt="Artibot Mascot" width="400" /> </p>

Overview

Most Claude Code plugins use simple sub-agent (Task()) delegation -- fire-and-forget with one-way result reporting. Artibot takes a fundamentally different approach by using Claude's native Agent Teams API as its core engine, enabling true team orchestration:

| Capability | Sub-Agent (Task) | Agent Teams (Artibot) | |------------|------------------|----------------------| | Communication | Return result to parent only | P2P bidirectional messaging (SendMessage) | | Task Management | Parent manages everything | Shared task list (TaskCreate/TaskList) | | Self-Assignment | Not possible | Teammates self-claim from task list | | Peer Communication | Not possible | Direct DM + broadcast | | Plan Approval | Not possible | plan_approval_response | | Lifecycle | One-shot | Create -> Work -> Shutdown -> Cleanup |

Key Features

• Cognitive Architecture -- Dual-process System 1/2 routing inspired by Kahneman's theory: fast intuitive responses for simple tasks, deep deliberative reasoning for complex ones
• Lifelong Learning -- GRPO-based batch learning from session outcomes with automatic knowledge transfer between System 1 and System 2 caches
• CTO-Led Orchestration -- orchestrator agent leads 28 specialist agents as a team coordinator (delegation mode: no direct coding)
• Intelligent Delegation -- Auto-selects Sub-Agent (simple) vs Agent Team (complex) based on cognitive complexity scoring
• 5 Orchestration Patterns -- Leader, Council, Swarm, Pipeline, Watchdog
• 8 DAG Playbooks -- Feature, Bugfix, Refactor, Security + Marketing Campaign, Marketing Audit, Content Launch, Competitive Analysis (DAG-based with parallel node execution, topological sort, and cycle detection)
• 78 Slash Commands -- /sc smart router, /save, /resume, /daily, /team, /orchestrate, /spawn, /implement, /visual-check, /sc playbook, /learning, /ultrareview, /ultraplan, /audit-claude-md, /export, and more
• 28 Specialized Agents -- Architecture, security, frontend, backend, testing, DevOps, marketing, SEO, analytics, and more (all on claude-opus-5)
• 113 Domain Skills -- 11 persona skills, 8 core skills, 16 language skills, 8 utility skills, 35 marketing skills, visual-validation, daily, team, session-worklog, vibe-coding, repo-benchmarking, git-worktree, dynamic-context-injection, claude-md-auditor, skill-authoring (all enhanced with Anthropic best-practice descriptions, workflow checklists, HITL v2 conversational checkpoints, output templates, and freedom levels)
• 8 Auto-Activating Rules -- DEV protocol, quality gates, agent coordination, config safety, frontend/backend/test patterns, clean state enforcement
• Guard Registry -- Centralized guard pipeline with registerGuard()/executeChain() API, 6 built-in guards extracted from hook scripts (75% code reduction)
• Loop Detection -- Circular buffer-based agent loop detection with fingerprint matching, automatic warn/block on repeated tool calls
• Clean State Enforcement -- TaskCompleted hook ensures lint+test verification at feature completion boundaries
• 57 Event Hook Registrations -- Cognitive routing, lifelong learning, session lifecycle, dangerous command blocking, auto-formatting, team tracking, loop detection, clean state checks, HTTP webhook notifications, git autopilot
• Advisory File Locking -- Spin-lock based file locking for concurrent hook state access, fail-open pattern prevents workflow blocking
• DEV Protocol -- Mandatory Decompose-Execute-Verify workflow with zero-skip policy for all code changes
• Vibe Coding Support -- Natural language request handling with read-first, verify-after, evidence-based completion
• Visual Validation Pipeline -- SSIM-based screenshot comparison, auto-fix suggestion, iterative correction loop via Playwright MCP
• Conversation-to-Memory -- Auto-extracts rules and decisions from user messages (Korean/English), injects into skills dynamically
• Project CLAUDE.md Seeding -- install.sh auto-generates project-level CLAUDE.md with Artibot methodology and DEV protocol
• Runtime Middleware Pipeline -- 11-stage middleware engine (lifecycle, router, memory, skills, tasks, subagents, guardrail, summarization, token-usage, checkpoint, cache-roi) for runtime context injection
• Output Design System -- 7 output styles with design tokens (tokens.md) and narrative template for consistent formatting
• Forked Context Skills -- All 113 skills run in isolated forked context for clean execution without cross-contamination
• HTTP Webhook Notifications -- Session events sent to Slack, Discord, or custom endpoints via configurable webhooks
• Cross-Platform Compatible -- Works with Gemini CLI, OpenAI Codex, and Cursor via platform adapters
• Zero Dependencies -- Pure Node.js built-in modules only (node:fs, node:path, node:os)

Get Started

Prerequisites

Agent Teams is auto-enabled by Artibot on first session start. No manual setup needed.

Installation

Recommended: native marketplace (no clone) — run inside any Claude Code session:

/plugin marketplace add Yoodaddy0311/artibot
/plugin install artibot@artibot

Claude Code fetches the plugin straight from GitHub into its plugin cache — no git clone, no shell script. Commands, agents, skills, and hooks load via ${CLAUDE_PLUGIN_ROOT}; updates come through /plugin marketplace update artibot. Native install namespaces every command under the artibot: prefix (call /artibot:save, /artibot:sc). Agent Teams auto-enables on first session start. To uninstall: /plugin uninstall artibot@artibot.

Native install does not deliver the 8 auto-activating rules (DEV Protocol, Quality Gates, agent-coordination, config-safety, clean-state, frontend/backend/test patterns). Claude Code's plugin schema has no rules field, so they load only when install.sh copies them to ~/.claude/rules/artibot/. If you want that automatic DEV-protocol / quality-gate enforcement, use the full install below. /theme and the themed statusline now auto-resolve both layouts (native install included); a native-only install may need /theme re-run after a major /plugin update. /update still assumes the flat layout and is being migrated. Commands, agents, skills, and hooks themselves work identically on both paths.

Full install: install.sh (flat commands, no prefix)

git clone https://github.com/Yoodaddy0311/artibot.git
cd artibot/plugins/artibot
bash install.sh          # macOS / Linux / Git Bash on Windows

This flat-copies agents and commands into ~/.claude/, so slash commands are called without a prefix (/save, /sc, /daily). It enables Agent Teams, wires the themed statusline, and seeds a conservative read-only permission allowlist (Read/Glob/Grep) into ~/.claude/settings.json, removing the repeated approval prompts new users hit on first run. To uninstall: bash install.sh uninstall.

Windows: run install.sh from Git Bash (Git for Windows), or use the native PowerShell installer: powershell -ExecutionPolicy Bypass -File install.ps1.

Command invocation differs by install method:

| Install method | Command form | Example | |---|---|---| | Native marketplace (/plugin install) | namespaced artibot: | /artibot:save, /artibot:sc | | install.sh / install.ps1 (flat) | flat (no prefix) | /save, /sc, /daily |

Requirements

• Claude Code CLI
• Node.js >= 18.0.0
• Agent Teams (auto-enabled by Artibot, or manually set CLAUDE_CODE_EXPERIMENTAL_AGENT_TEAMS=1)

Cross-Platform Installation

Artibot works beyond Claude Code. Built-in adapters auto-convert agents, skills, and commands for each platform.

Replace <your-project> with the root directory path of the target project.

| Platform | Compatibility | Agent Teams | Sub-Agent | Skills | |----------|:------------:|:-----------:|:---------:|:------:| | Claude Code | 10/10 | ✅ | ✅ | ✅ | | Gemini CLI | 9/10 | ❌ | ✅ | ✅ | | Codex CLI | 8/10 | ❌ | ✅ | ✅ | | Antigravity | 8/10 | ❌ | ✅ | ✅ | | Cursor IDE | 6/10 | ❌ | ⚠️ | ✅ |

Gemini CLI

git clone https://github.com/Yoodaddy0311/artibot.git
cd artibot
node --input-type=module -e "import { exportForGemini } from './plugins/artibot/lib/adapters/skill-export.js'; const r = await exportForGemini({ pluginRoot: './plugins/artibot' }); console.log(r.summary);"

OpenAI Codex CLI

git clone https://github.com/Yoodaddy0311/artibot.git
cd artibot
node --input-type=module -e "import { exportForCodex } from './plugins/artibot/lib/adapters/skill-export.js'; const r = await exportForCodex({ pluginRoot: './plugins/artibot' }); console.log(r.summary);"

Google Antigravity

git clone https://github.com/Yoodaddy0311/artibot.git
cd artibot
node --input-type=module -e "import { exportForGemini } from './plugins/artibot/lib/adapters/skill-export.js'; const r = await exportForGemini({ pluginRoot: './plugins/artibot' }); console.log(r.summary);"

Note: Agent Teams API is exclusive to Claude Code. Other platforms automatical