multi-agent-cli-bridge

C runtime + Python ↔ GUI bridge for a multi-agent CLI assistant. The runtime layer I authored as part of a private team-owned multi-agent desktop assistant project, extracted as a portfolio demonstration. Runs on top of two LoRA adapters I trained (jalva182/cli-agent-model, jalva182/cli-agent-model-gpu1).

Restricted-use repository. Source-available for review. See NOTICE.md — no license is granted; all rights reserved.

What this is

A multi-agent CLI assistant has three layers in roughly the following order:

1. An event source that observes the user's system (process events, file events, hotkeys).
2. A routing layer that decides which "expert" agent should handle a given query, given the event context.
3. A runtime/bridge that loads a fine-tuned LLM and surfaces responses through a desktop UI.

This repository is layers 2 and 3 — the routing layer (master_router.py), the C runtime + Python bridge (py_bridge.c, system_monitor.c, hotkey_daemon.c, c_bridge.py), and two Tkinter GUIs that drive the assistant.

The event source for layer 1 lives in the team-private upstream and is not included here. A minimal SystemEvent dataclass (src/c_core/system_event.py) stands in for it so the routing layer can be exercised standalone.

Repo layout

.
├── src/
│   ├── c_core/
│   │   ├── py_bridge.c        # Python <-> C bridge (CPython API)
│   │   ├── system_monitor.c   # cross-platform system stats (Windows API + sys/...)
│   │   ├── hotkey_daemon.c    # global hotkey listener stub
│   │   ├── c_bridge.py        # Python-side wrapper over the C runtime
│   │   ├── master_router.py   # routes queries + events to expert agents
│   │   ├── system_event.py    # minimal standalone SystemEvent (substitute your own)
│   │   ├── ai_gui.py          # Tkinter chat GUI
│   │   ├── gui_simple.py      # simpler launcher GUI variant
│   │   └── __init__.py
│   └── scripts/
│       ├── setup_env.sh
│       └── setup_env.bat
├── BUILD_INSTRUCTIONS.md
├── CMakeLists.txt
├── Makefile
├── build.sh
├── setup.py
├── pyproject.toml
├── requirements.txt
├── setup_project.bat
├── verify_installation.ps1
├── test_run.py                # smoke test
├── NOTICE.md
└── README.md

Trained models (the LLMs this layer runs)

The bridge is designed to run on top of LoRA adapters fine-tuned with GRPO via Unsloth + TRL on unsloth/llama-3-8b-Instruct. Both adapters and the training code are public.

Model	Base	Method	Link
jalva182/cli-agent-model	unsloth/llama-3-8b-Instruct	GRPO + LoRA (TRL via Unsloth)	https://huggingface.co/jalva182/cli-agent-model
jalva182/cli-agent-model-gpu1	unsloth/llama-3-8b-Instruct	GRPO + LoRA (TRL via Unsloth)	https://huggingface.co/jalva182/cli-agent-model-gpu1

Training pipeline: Alvarez-Jose/unsloth-grpo-project (extending @Carson1829's GRPO training section).

Quick start

git clone https://github.com/Alvarez-Jose/multi-agent-cli-bridge
cd multi-agent-cli-bridge

# Python deps
pip install -r requirements.txt

# Build the C runtime (Windows / WSL — see BUILD_INSTRUCTIONS.md for full instructions)
bash build.sh

# Smoke-test the import surface
python test_run.py

# Launch the GUI
python -m c_core.ai_gui

For full build details, including CMake and the verification PowerShell script, see BUILD_INSTRUCTIONS.md.

What's not included

• The expert-agent implementations (DebuggingExpert, FileOperationsExpert, etc.) — owned by the Visualtaggy team, private.
• The ETW event-stream wiring that produces real SystemEvent instances in production — also private. The included system_event.py is a minimal standalone replacement for portfolio demonstration.
• The full multi-agent orchestration, agent-handoff logic, and conversation memory layer — also private.

If you're reviewing this as a recruiter or collaborator and want to see how these pieces fit together end-to-end, contact me directly.

Credits

• Broader multi-agent system this layer plugs into: Visualtaggy team (private upstream).
• GRPO training section that produced the LoRA adapters extended in my fork at Alvarez-Jose/unsloth-grpo-project.

License

See NOTICE.md. All rights reserved; no license is granted. Restricted-use; source-available for review.

---

Author: Antonio Alvarez Maciel · M.S. NLP, UC Santa Cruz · LinkedIn · Email