datapath_agt (Open Source MVP)

A lightweight multi-agent framework for hardware RTL generation + verification:

• Decomposer Agent: breaks a high-level requirement into a module tree/contracts
• RTL Coder: generates Verilog RTL for leaf modules
• TB Generator: generates a top-level testbench based on the real DUT ports
• Auto-Rework Loop: runs iverilog and, on compile failures, invokes a Rework Agent to overwrite the RTL bundle and retry (up to 5 attempts)

Quick Start

1) Install

pip install -r requirements.txt

2) Toolchain (required for simulation + rework)

Install Icarus Verilog so iverilog and vvp are available in your PATH.

3) Run (recommended: mock mode first)

python -m orchestrator.run_mvp --mock --user-goal-file user_goal.txt --no-progress

Outputs are written to ./output/<timestamp>/:

• hdl_bundle.v (generated RTL bundle)
• tb_top.v (generated testbench)
• simulation.log (build/run output)
• sim.vvp (iverilog output)

4) Run with a real LLM

Create a local .env (do not commit it) based on .env.example.

Defaults:

• If OPENAI_API_KEY is set: uses OpenAI-compatible Chat Completions with LLM_MODEL=gpt-4o.
• If DEEPSEEK_API_KEY is set (and OpenAI key is empty): uses DeepSeek OpenAI-compatible API with LLM_MODEL=deepseek-chat.
• If both are set: prefers OpenAI.

You can override via LLM_PROVIDER, LLM_API_URL, LLM_MODEL, LLM_TEMPERATURE (see orchestrator/config.py).

Project Layout

• orchestrator/ core runtime
• prompts/ prompt registry (required)
• user_goal.txt example goal you can edit