🤖 DV Agentic System

⚡ Why this exists

Hardware verification produces huge logs, deep coverage questions, and repeated reasoning across sessions. Raw LLM access is not enough: you need structured prompts, team and IP context, and guardrails that match how DV teams actually work.

This repository is the shared Python package that provides that scaffolding: agent orchestration, prompt loading, configuration wiring, and integration points for simulators and coverage tools. Organization-specific rules and VIP catalogs live outside this repo (see Architecture).

🎯 What you get here

• A multi-agent workflow oriented to UVM verification (orchestrator plus specialized agents).
• A three-layer split between generic code (this repo), org profiles, and per-project .agent/ configuration.
• Documentation for architecture, prompts, APIs, and CLI — see the documentation home.

Design constraint (policy): agents follow a testbench-only stance toward RTL: they do not treat RTL sources as a writable surface; hardware intent is driven by your verification artifacts (e.g. specification inputs and plans you configure), not by editing DUT RTL through this framework.

For current capabilities, agent list, and roadmap, use the docs and changelog rather than this file:

• Documentation home
• Agents
• CLI

🏗️ Architecture

┌─────────────────────────────────────────────────────┐
│  Layer 1 · Shared package (this repository)         │
│  Agent logic, adapters, base prompts — no org IP    │
├─────────────────────────────────────────────────────┤
│  Layer 2 · Team profile repository (yours, private) │
│  Protocol rules, VIP index, prompt patches          │
├─────────────────────────────────────────────────────┤
│  Layer 3 · Project implant ({project}/.agent/)      │
│  project.yaml connects Layer 1 and Layer 2          │
└─────────────────────────────────────────────────────┘

Details: Structure & design, System overview.

🧠 LLM Wiki (optional knowledge layer)

Some workflows persist verification notes and summaries in a Git-versioned Markdown wiki under the project’s .agent/ tree so context can carry across sessions.
The pattern (persistent markdown between you and raw sources, maintained by an agent rather than re-derived on every question like chunk-RAG alone) follows the public “LLM Wiki” idea file by Andrej Karpathy (GitHub Gist, April 2026).
This repository’s layout, tooling, and agent hooks are our instantiation; use the gist for the original motivation and the docs below for what is implemented here.

• LLM Wiki — architecture and usage

📦 Requirements

• Python: see requires-python in pyproject.toml (source of truth).
• Package manager: this repo is set up for uv.

🚀 Quick start (contributors & local development)

git clone https://github.com/anlit75/dv-agentic-system.git
cd dv-agentic-system
uv sync

Optional: activate the virtual environment (source .venv/bin/activate on Unix, .venv\Scripts\activate on Windows).

uv run pre-commit install
uv run pre-commit run --all-files

✨ Option — setup via an AI assistant (copy-paste prompt)

Use the block below as a single message to your AI assistant.

📋 Copy prompt for Cursor / Claude Code / Copilot Chat (etc.)

You are helping me set up the local dev environment for the dv-agentic-system repo (Python, uv).

Goals:
1) Install dependencies: run `uv sync` from the repo root.
2) (Optional) Install git hooks: `uv run pre-commit install`, then `uv run pre-commit run --all-files`.
3) Agent installer: run `uv run python -m dv_agentic.cli.install_agents --help`, then run the installer with the `--target` value that matches my tools (choices are printed by --help). If `./scripts/install-agents.sh` exists, you may use it instead after reading `./scripts/install-agents.sh --help`.
4) Report any failures with the exact command and stderr.

Constraints:
- Do not invent CLI flags; only use what --help lists.
- On Windows, if symlink creation fails for mirrored assets, note Developer Mode / admin requirements and continue if the installer falls back to copies.

When done, summarize what ran and what I should run next manually if anything needs human approval.

🤖 Install agent prompt bundles into a project

The installer writes agent markdown and mirrored assets into paths expected by supported AI coding tools. Which paths are populated depends on the flags you pass; do not assume defaults without reading help output once:

uv run python -m dv_agentic.cli.install_agents --help

On Unix you can also use ./scripts/install-agents.sh --help if present.

📚 Documentation site (local)

uv sync --all-groups
uv run mkdocs serve

🔒 Air-gapped install

Use the offline install guide: bundle layout, release assets vs. offline-download.sh, and offline-install.sh behavior.

🔬 Research note

Design choices are informed by public benchmarks in the verification / LLM literature. One widely cited dataset is CVDP (Pinckney et al., 2025) — arXiv:2506.14074.

📄 License

MIT.