Skip to content

Repository files navigation

Multi-Platform Agentic AI Engineering Operating System & Sub-Worker Delegation Engine

Architecture: Two-Tier Multi-Agent Standard: agentskills.io MCP: FastMCP v1.0 License: MIT Platforms: Linux | macOS | Windows

A deterministic, cross-platform engineering suite and operating framework for leading AI coding assistants (OpenAI Codex, Antigravity / Gemini CLI, Claude Code, and Modern Cursor IDE (.mdc)) across Linux, macOS, and Windows.

The bundle unifies 16 ironclad operational rules (Prime Directives), 36 modular skills (agentskills.io standard), and an autonomous Two-Tier Cognitive Architecture powered by a custom FastMCP Sub-Worker Delegation Bridge (chinese-worker) and Aider Headless Engine. Expensive frontier models (Gemini 3.7 Pro, Claude 3.7 Sonnet) act as high-level architects and orchestrators, delegating repetitive bulk coding tasks (scaffolding 2500+ LOC, TDD unit test generation, JSDoc/docstrings, strict type migrations) to 100% free, mega-context models (MiniMax M3 1M Context, NVIDIA Nemotron 550B MoE, Zhipu GLM-5.2) in sandboxed Git Worktrees with Zero Context Bleed and Zero Token Cost ($0.00).


⚡ 1-Prompt Autonomous AI Installation

If you are using an AI coding assistant (Claude Code, OpenAI Codex, Cursor Composer, Gemini CLI / Antigravity, Cline, Roo Code, OpenCode, Aider), simply paste this prompt:

English:

Clone https://github.com/Gzyms69/agent-setup-bundle.git and install the full AI engineering operating system for me following AGENTS.md in the repo.

Polski:

Sklonuj https://github.com/Gzyms69/agent-setup-bundle.git i zainstaluj całe środowisko inżynieryjne według instrukcji w AGENTS.md.

Your AI assistant will read AGENTS.md, auto-detect your operating system (Windows, macOS, or Linux), execute the native installer, validate suite integrity, and immediately adopt the Senior AI Pair Programmer persona.


1. Executive Summary & Two-Tier Cognitive Architecture

Traditional single-model AI pair programming ("Just write this feature") fails on complex production codebases due to four fundamental problems:

  1. Economic Inefficiency: Burning expensive reasoning tokens on repetitive boilerplate, large mock suites, or syntax formatting.
  2. Context Window Degradation: Dumping thousands of lines of raw logs and intermediate diffs into the primary reasoning context.
  3. Simulation & Mocking Traps: Writing fake stubs or untested boilerplate without running verification suites.
  4. Monolithic Spaghetti Code: Coupling business logic to UI frameworks and creating unmaintainable "god files".

agent-setup-bundle solves this through a Two-Tier Cognitive Architecture:

flowchart TD
    subgraph Tier1_Brain ["Tier 1: High-Reasoning Brain & Orchestrator ($$)"]
        Gemini["Gemini 3.7 Pro / Antigravity CLI"]
        Claude["Claude 3.7 Sonnet / Claude Code"]
        Cursor["Cursor IDE / Composer"]
        Codex["OpenAI Codex CLI"]
        SpecGate["Spec-Driven Development & 5-Axis Code Review Gate"]
        Gemini & Claude & Cursor & Codex --> SpecGate
    end

    subgraph Tier2_FastMCP ["Tier 2: FastMCP Sub-Worker Bridge (chinese-worker) ($0.00)"]
        Router["Intelligent Task Router (Keyword & Task Affinity)"]
        SkillsInj["Dynamic Skill Injector (--read ~/.agents/skills/*/SKILL.md)"]
        WorktreeMgr["Git Worktree Sandbox (.git/worktrees_active/<task-id>)"]
        AiderEngine["Aider Headless Engine (Diff Mode)"]
        SelfHealing["Self-Healing Quality Loop (--auto-test)"]
        
        SpecGate -->|MCP Tool: worker_run_task / worker_generate_tests| Router
        Router --> WorktreeMgr --> AiderEngine
        Router --> SkillsInj --> AiderEngine
        AiderEngine --> SelfHealing
    end

    subgraph Model_Pool ["Free Tier & High-Throughput Model Pool"]
        M3["MiniMax M3 (1M Ctx, 65k Out) -> Mega Scaffolding & Fullstack"]
        N550["NVIDIA Nemotron 550B MoE -> Low-Level, C++, ASM & Math"]
        GLM5["Zhipu GLM-5.2 (LiveCodeBench 74-85%) -> Bugfix & Refactor"]
        NLight["Nemotron 3.5 Lightning -> Rapid TDD Unit Tests"]
        GLM4["Zhipu GLM-4-Flash PAAS -> 100% Guaranteed Direct Fallback"]
        
        Router --> Model_Pool
        Model_Pool --> AiderEngine
    end

    SelfHealing -->|Zero Context Bleed: 3-line Structured Report| SpecGate
    SpecGate -->|Inspection & Approval| Merge["worker_merge_branch"]
Loading
  1. Tier 1 (Brain & Orchestrator): Frontier models (Gemini 3.7, Claude 3.7) plan architecture, conduct Spec-Driven Development, and enforce the 5-Axis Code Review Gate.
  2. Tier 2 (Grunt Worker & Token Factory): FastMCP (chinese-worker) delegates bulk implementation tasks to free, specialized models operating inside sandboxed Git Worktrees with an automated Self-Healing Quality Loop (--auto-test).

2. Universal 4-Phase Pre-Flight Skill Gate

Before executing file discovery (grep_search, find_by_name, view_file) or making any code changes, agents operating under this system MUST evaluate and activate skills through the 4-Phase Pre-Flight Skill Gate:

flowchart LR
    P0["Phase 0: Cartography Gate"] --> P1["Phase 1: Planning & Orchestration"]
    P1 --> P2["Phase 2: Domain Specialists"]
    P2 --> P3["Phase 3: QA & Review Gate"]
Loading

Complete 36-Skill Matrix (~/.agents/skills/):

Phase Skill Name Trigger & Responsibility
Phase 0 skill-codebase-onboarding Mandatory first step for exploring or onboarding any unmapped repository.
Phase 0 spec-miner Reverse-engineering legacy, undocumented, or poorly structured codebases.
Phase 1 spec-driven-development Tasks >15 min, >3 files, architectural decisions, or /plan invocation.
Phase 1 skill-context-engineering Attention budget curation, log offloading to ./scratch/, context compaction.
Phase 1 skill-master-orchestrator Multi-agent swarm coordination, Task DAG decomposition, barrier synchronization.
Phase 1 skill-monorepo-architect Monorepo structure management (PNPM, Turborepo, UV workspaces).
Phase 1 skill-plugin-architecture Extensible microkernel systems, dynamic toolkits, plugin discovery.
Phase 1 skill-web-architecture Full-stack web architectural standards, module boundaries, API contracts.
Phase 2 skill-frontend-architect Next.js 15+ App Router, RSC, Client Island boundaries, WCAG 2.1/2.2 AA.
Phase 2 skill-design-engineering Motion animations (motion.dev), CSS Subgrid, Container Queries, 21st.dev UI.
Phase 2 skill-creative-design Art direction, aesthetics, Fontjoy typography math, OKLCH color physics.
Phase 2 skill-backend-architect Backend architecture, database schemas, API contracts, zero-downtime migrations.
Phase 2 skill-mcp-builder Model Context Protocol server development (FastMCP, TypeScript SDK, stdio/SSE).
Phase 2 skill-low-level-programming C/C++, Rust, Assembly, byte manipulation, memory layout, bitwise arithmetic.
Phase 2 c-cpp-systems Low-level C/C++ memory safety, pointers, manual RAII, ASan/UBSan sanitizers.
Phase 2 wasm-emscripten C/C++ to WebAssembly compilation, Emscripten runtime bridging, HEAP views.
Phase 2 retro-emulation-engineering Retro emulator architecture, hardware coprocessor simulation (CPU/RSP/RDP).
Phase 2 skill-emulator-wasm WebAssembly retro emulators, WebGL rendering, Web Audio sync, save states.
Phase 2 skill-ai-ml LLM integrations (Gemini, OpenAI, Anthropic), RAG pipelines, vector DBs.
Phase 2 skill-data-science Data science workflows, exploratory data analysis (EDA), ingestion pipelines.
Phase 2 skill-data-analysis Statistical methodology, hypothesis testing, anomaly detection, claim validation.
Phase 2 skill-graph-analytics Graph databases (Neo4j), Cypher queries, topology analysis, Graph Data Science.
Phase 2 skill-graphics-webgl 2D/3D graphics, Three.js, WebGL shader optimization, Canvas rendering.
Phase 2 skill-stealth-scraping Anti-bot evasion, stealth automation, TLS/JA3 fingerprints, reverse API engineering.
Phase 2 skill-osint-engineering OSINT intelligence pipelines, standardized entity graphs, pivoting engines.
Phase 2 skill-system-diagnostics Hardware/OS/kernel diagnostics, log analysis, SRE root-cause debugging.
Phase 2 skill-devops-cloud Docker containerization, CI/CD pipelines, Cloud Run checklists, Kubernetes.
Phase 2 skill-research Academic and technical literature research with multi-source verification.
Phase 2 skill-resume-tailor AI developer resume/CV architect (Google XYZ formula, Harvard Tech standard).
Phase 2 marketing-copywriting Conversion-focused copywriting, value propositions, CTA engineering.
Phase 2 avoid-ai-writing Strict audit and rewriting protocol eliminating AI writing clichés and fluff.
Phase 2 seo-optimization-and-audit SEO audit, metadata, head tags, Core Web Vitals optimization.
Phase 2 skill-web-performance Universal web performance engineering, Lighthouse 100/100, runtime tracing.
Phase 3 skill-qa-engineer Mandatory Phase 3 QA Gate, TDD Red-Green discipline, TypeScript Safety Gate.
Phase 3 skill-code-review Mandatory 5-axis review (Correctness, Readability, Architecture, Security, Perf).
Phase 3 doubt-driven-development Adversarial verification gate challenging false confidence before assertions.

3. 16 Operational Rules (The Laws of Robotics)

All operational rules reside in ~/.agents/rules/ and are enforced across all platform manifests:

  1. zero-speculation.md: Total ban on guessing hardware specs, package versions, API endpoints, or error causes. Verify via live commands or web search.
  2. command-verification.md: Mandatory verification of CLI tool outcomes before proceeding.
  3. env-integrity.md: Pre-flight environment audit before modifying codebase configuration.
  4. error-triage.md: Strict diagnostic triage sequence: Docs -> Web -> Source Code.
  5. full-log-reporting.md: Zero truncated logs when diagnosing failures.
  6. problem-isolation.md: Surgical problem isolation without collateral workspace mutation.
  7. subagent-economy.md: Subagent model routing economy (flash_lite -> flash -> pro) and Chinese worker delegation.
  8. system-identity.md: Real-time hardware identity and OS verification template.
  9. systemic-excellence.md: Prohibition of symptomatic patches, workarounds, or defensive masking.
  10. context-engineering.md: Strict 100-line / 5 KB offloading to ./scratch/ and Attention U-Curve protection.
  11. modular-architecture.md: Clean/Hexagonal architecture boundaries and anti-god-file constraints.
  12. planning-and-document-integrity.md: Stateful 3-state planning machine, Iteration Delta, and Discovered Facts Lock.
  13. session-handoff.md: Lossless session transition via NEXT_SESSION_PLAN.md and clean SSOT handoff prompts.
  14. skill-orchestration.md: Universal 4-Phase Pre-Flight Skill Gate activation protocol.
  15. mcp-master-playbook.md: Standardized Model Context Protocol tool invocation guidelines.
  16. mempalace-discovery.md: Knowledge graph querying and memory retrieval protocol.

4. Autonomous FastMCP Sub-Worker Bridge (chinese-worker)

The bundle includes a native, high-throughput Model Context Protocol server implemented in Python using FastMCP (scripts/worker_mcp.py).

Native FastMCP Tools:

Tool Name Parameters Description
worker_run_task instruction, editable_files, readonly_files, skills, task_type, profile, auto_test, test_cmd, use_worktree Executes an autonomous coding task in a dedicated Git Worktree with dynamic skill injection and self-healing test loop.
worker_generate_tests target_file, test_file, test_framework, skills, profile Generates comprehensive TDD unit tests (pytest, vitest, jest, cargo) with edge case mocking.
worker_generate_docs target_files, doc_type, profile Generates JSDoc, docstrings, or markdown guides preserving exact code functionality.
worker_batch_refactor target_files, instruction, readonly_files, skills, profile Executes mass refactoring or strict type safety upgrades across multiple files.
worker_continue_task task_id, feedback Continues refining changes within an existing active worktree sandbox.
worker_get_diff task_id Returns the clean unified git diff generated by the worker for inspection.
worker_merge_branch task_id, target_branch Merges the verified task worktree into the main working tree and cleans up.
worker_discard_branch task_id Deletes and cleans up a rejected task worktree sandbox.
worker_status (none) Lists all active worker worktrees and recent task logs.

Intelligent Model Affinity & Task Routing:

{
  "task_affinity": {
    "scaffold": "minimax-m3",
    "fullstack": "minimax-m3",
    "low_level": "nemotron-550b",
    "binary": "nemotron-550b",
    "algorithms": "nemotron-550b",
    "bugfix": "glm-5.2",
    "refactor": "glm-5.2",
    "tests": "nemotron-lightning",
    "unit_tests": "nemotron-lightning",
    "docs": "glm-5.2",
    "fast": "cohere-code"
  }
}
  1. MiniMax M3 Free (openrouter/minimax/minimax-m3:free): Tier 0 Chinese Flagship with 1,048,576 Context and 65,536 Max Output Tokens. Best for fullstack scaffolding and large multi-file codebases.
  2. NVIDIA Nemotron 3 Ultra 550B MoE (openrouter/nvidia/nemotron-3-ultra-550b-a55b:free): 550 Billion parameter MoE with 1M context. Specialized in low-level systems (C/C++, Rust, Assembly, byte manipulation, and mathematical algorithms).
  3. Zhipu GLM-5.2 Free (openrouter/z-ai/glm-5.2:free): Frontier coding model with LiveCodeBench 74-85% and SWE-bench 68.2%. Specialized in bugfixes, refactoring, and diff generation.
  4. NVIDIA Nemotron 3.5 Lightning (openrouter/nvidia/nemotron-3.5-lightning:free): Ultra-fast MoE for rapid TDD unit test creation.
  5. Zhipu GLM-4-Flash PAAS (openai/glm-4-flash): Direct BigModel PAAS integration serving as 100% guaranteed fallback if OpenRouter free endpoints hit temporary rate limits.

Developer CLI Companion (worker):

The installer provisions a terminal CLI symlink at ~/.local/bin/worker:

# Diagnostic health check (validates packages, profiles, API keys)
worker check

# Interactive coding session with MiniMax M3
worker chat minimax-m3 --skills skill-frontend-architect src/App.tsx

# Batch instruction execution with automatic model routing
worker run "Refactor database queries to use parameterized prepared statements" --skills skill-backend-architect -f src/db.ts

5. Model Context Protocol (MCP) Multi-Server Matrix

The bundle provisions unified MCP server configurations across Gemini CLI (config/mcp_config.json), Cursor (config/cursor_mcp.json), and Claude Code:

MCP Server Provider / Package Purpose
chinese-worker scripts/worker_mcp.py (FastMCP) High-throughput autonomous sub-worker delegation engine ($0.00).
github @modelcontextprotocol/server-github Remote GitHub API operations (PRs, issues, code search, reviews).
chrome-devtools chrome-devtools-mcp@latest Headless Chrome browser automation and DOM inspection.
puppeteer @modelcontextprotocol/server-puppeteer End-to-end web testing and screenshot capture.
lighthouse-mcp @danielsogl/lighthouse-mcp Web performance, Core Web Vitals, and accessibility audits.
postgres @modelcontextprotocol/server-postgres PostgreSQL schema introspection and query analysis.
sqlite @modelcontextprotocol/server-sqlite Local SQLite database inspection.
docker @modelcontextprotocol/server-docker Container lifecycle management and log inspection.
firecrawl firecrawl-mcp Web scraping, crawling, and clean Markdown extraction.
ast-grep @ast-grep/mcp Structural AST search and code pattern matching.
mempalace mempalace Long-term memory palace, AAAK knowledge graph and diary storage.

6. Multi-Platform Installation Guide

Option 1: Native Shell Installers

Linux & macOS (Bash):

git clone https://github.com/Gzyms69/agent-setup-bundle.git
cd agent-setup-bundle
chmod +x install.sh
./install.sh --all

Windows (PowerShell 5.1 / 7+):

git clone https://github.com/Gzyms69/agent-setup-bundle.git
cd agent-setup-bundle
powershell -ExecutionPolicy Bypass -File .\install.ps1 -All

Universal Python 3 Installer (All OSes):

python3 install.py --all

Option 2: Selective Installation Flags

  • --codex / -Codex: Install only OpenAI Codex environment (~/.codex/).
  • --gemini / -Gemini: Install only Antigravity / Gemini CLI environment (~/.gemini/).
  • --claude / -Claude: Install only Claude Code environment (~/.claude/).
  • --cursor / -Cursor: Install only Cursor IDE rules (~/.cursor/rules/).

7. Quality Assurance & Automated Testing

Every component in this repository is strictly validated by automated test suites before deployment:

# 1. Run the master suite integrity validator (16 rules, 36 skills, 4 platforms, worker configs)
python3 scripts/validate_suite.py

# 2. Run the Worker MCP & CLI unit test suite
python3 scripts/tests/test_worker.py

# 3. Run the Sub-Worker environment diagnostics
python3 scripts/worker_cli.py check

8. Directory Topography

agent-setup-bundle/
├── AGENTS.md                        # Master repository blueprint & AI installer instructions
├── README.md                        # Master technical documentation & cross-platform guide
├── CAREER_KNOWLEDGE_BANK.md         # Master SSOT for career portfolios, metrics & STAR+R cases
├── PROMPT_FOR_AI.md                 # Universal bootstrap prompts
├── llms.txt                         # Semantic summary for web-enabled LLM agents
├── install.sh                       # Native Bash installer (Linux / macOS)
├── install.ps1                      # Native PowerShell installer (Windows)
├── install.py                       # Universal Python 3 installer (All OSes)
├── core/                            # Platform manifests (CODEX.md, GEMINI.md, CLAUDE.md, cursor)
├── rules/                           # 16 Universal Operational Rules (~/.agents/rules/)
├── skills/                          # 36 Modular Skills (~/.agents/skills/)
├── templates/
│   ├── AGENTS.md                    # Project-level starter template
│   ├── CONVENTIONS.md               # Universal coding conventions for sub-workers
│   └── .aider.conf.yml.template     # Universal Aider configuration template
├── config/
│   ├── worker_profiles.json         # Sub-worker model routing profiles & context bounds
│   ├── .aider.model.settings.yml    # Aider model behavioral settings & diff formats
│   ├── .aider.model.metadata.json   # Aider token limit overrides (1M context / 65k output)
│   ├── mcp_config.json              # Gemini CLI MCP configuration template
│   ├── settings.json                # Gemini CLI general settings
│   ├── codex_config.toml            # OpenAI Codex configuration template
│   └── cursor_mcp.json              # Cursor IDE MCP configuration template
├── policies/                        # MCP tool planning policies
└── scripts/
    ├── worker_mcp.py                # FastMCP server for autonomous sub-worker delegation
    ├── worker_cli.py                # Developer CLI companion (worker)
    ├── tests/
    │   └── test_worker.py           # Unit test suite for worker ecosystem
    └── validate_suite.py            # Quality assurance test suite

9. License

MIT License. Designed and maintained by Gzymson for autonomous, deterministic AI pair programming.

About

Multi-platform AI Engineering Operating System: 16 core rules (PRAR, Context Engineering, /plan protocol), 36 skills (agentskills.io), and MCP integration for Codex, Gemini CLI, Claude Code & Cursor

Topics

Resources

Stars

0 stars

Watchers

0 watching

Forks

Releases

Packages

Contributors

Languages