chipweaver/veripower

veripower

An open-source agent flow from natural language spec through Verilog RTL and UVM verification to front-end signoff on commercial EDA tools

未分類ai-agentschip-designclaude-code

VeriPower

Chip front-end design and verification with coding agents

English | 中文


VeriPower helps coding agents develop chip front-end designs from original requirements through implementation, verification and power, performance and area (PPA) evaluation. Its goal is a complete delivery in which the implementation and its checking artifacts jointly meet the task requirements.

Implemented as an open-source plugin for agent harnesses such as Claude Code, OpenCode, DeepSeek Harness and Codex, VeriPower combines professional guidance from domain-specific skills with a workflow engine. Design and checking evolve together through continuing feedback and coordinated revisions under shared requirements.

Design and checking evolve under shared requirements; an orchestrating agent and workflow engine coordinate task execution across iterations
Development method and execution mechanisms · View full size

What it provides

  • Design reasoning and early feedback. Functional requirements and PPA targets guide specification and RTL design. Top-down requirements refinement and bottom-up feasibility analysis inform structural choices and resource trade-offs. Verification planning and independent reviews provide early feedback to assess and refine those choices as the design evolves.
  • Requirements-driven verification. Original requirements and independent references guide testpoints, reference models and assessment criteria. Functional simulation, lint/CDC, synthesis, timing and power analysis check behavior and engineering targets. Independent reviews assess whether stimuli, observations and checking logic can detect violations and trigger test failure.
  • Diagnosis and coordinated revision. Executors combine diagnosis with change-impact analysis to identify the task responsible for a repair and assess its effects on other requirements and artifact consistency. The responsible task performs the revision in coordination with affected work.

The workflow engine connects these activities across iterations. It coordinates execution through artifact dependencies and links checking conclusions to artifact versions, identifying work that needs reassessment while retaining passing results that still apply. Artifacts and run records persist in the module workspace, allowing development to continue across sessions.

Deliverables include specifications, RTL, UVM testbenches, constraints, synthesized netlists, and verification and analysis reports.

Requirements

The agent's execution environment needs:

  • Python 3.10 or later, with the packages in requirements.txt.
  • For the full flow, Synopsys SpyGlass, VCS with UVM, Verdi waveform tools, Design Compiler and PrimeTime, with the corresponding licenses and cell libraries. See EDA Tool Environment for configuration.

You can check tool availability and license readiness with the env-precheck skill in a separate session before starting development.

Quickstart

  1. Install and enable VeriPower using your platform's guide: Claude Code, OpenCode, DeepSeek Harness or Codex. Start a session and ask the agent to list the VeriPower skills to confirm that the plugin loaded.

  2. Prepare the original requirements at {module_dir}/intent/brainstorm.md and copy referenced material into {module_dir}/intent/. Replace {module_dir} with the absolute path to your module workspace. An existing suitable specification can be used directly. If the requirements need discussion, use the brainstorm skill in a separate session and review the resulting document before starting the flow.

  3. Start a new session with VeriPower enabled and ask:

    Run the design-flow skill for {module_dir}.

The agent proceeds within your authorization and brings back questions that need your input. Published artifacts are under Design/ and Verification/ in the module workspace.

Documentation

  • User Guide: prepare inputs, follow progress, review results, revise requirements and resume interrupted work.
  • Architecture: development method, task responsibilities and coordination across iterations.
  • Contributing: development principles and validation.

Version: v0.2.10 · MIT License · Issues