Multi-CLI Multi-Agent Auto-Tuning Framework
Multiple AI coding CLIs coordinate via tmux — no external orchestration framework required. Pluggable strategies adapt it to various tasks.
Without a custom requirement definition, VibeCodeHPC runs this HPC auto-tuning workflow out of the box. Custom strategies can replace or extend it.
- Hierarchical Multi-Agent: PM → SE ↔ PG × N → CD
- Pluggable Strategies: HPC parallelization, local LLM deployment, GPU optimization — add your own
- Evolutionary Exploration: Flat directory structure for parallel search
- tmux IPC: Inter-agent communication with no special runtime
- 4-tier SOTA Tracking: Local → Family → Hardware → Project
User prepares 3 items:
requirement_definition.md— edit from template, or ask PM to create it interactively_remote_info/— site-specific info (details)BaseCode/— your code to optimize
Everything else is created by agents at runtime.
📂 VibeCodeHPC/ 🤖 PM ⬛
├── 📝 requirement_definition.md # ← User edits
├── 📁 _remote_info/ # ← User provides
├── 📁 BaseCode/ # ← User provides
│
├── 📁 User-shared/ # → Results here
├── 📂 Agent-shared/
│ ├── 📁 skills/ # Knowledge + scripts
│ └── 📁 logs/ # Agent communication history
│
├── 📄 CLAUDE.md # Common rules
├── 📁 instructions/ # PM, SE, PG, CD
├── 📁 vibecodehpc/ # Framework
Runtime directories (created by PM)
├── 📄 directory_pane_map.md
├── 📁 GitHub/ 🤖 CD ⬜
│
└── 📂 Flow/TypeII/single-node/ 🤖 SE1 🟦
├── 📄 hardware_info.md
├── 📂 gcc/
│ └── 📂 OpenMP/ 🤖 PG1.1 🟩
│ └── 📄 ChangeLog.md
├── 📂 nvidia/
│ ├── 📁 CUDA/ 🤖 PG1.2 🟧
│ └── 📁 OpenACC/ 🤖 PG1.3 🟪
└── 📂 intel/
└── 📁 OpenMP/ 🤖 PG1.4 🟥
Layout is determined by PM based on the requirement definition. Compiler/strategy hierarchy is configurable.
Claude Code, Codex CLI, Cline CLI, Gemini CLI, OpenCode, vibe-local, Qwen Code, Kimi Code CLI
- 📄 arXiv (v3) — iWAPT 2026
- 🎬 Demo video (EN subtitles)




