NovaCode — AI skill for Claude Code


NOVA : Agentic Coding Tool

[](https://github.com/Babitdor/NovaCode) [](https://opensource.org/licenses/MIT) [](https://www.python.org/)

An open-source, terminal-based AI coding assistant built on LangGraph and the `deepagents` framework. NOVA runs entirely in your terminal with both a Rich console REPL and a Textual TUI — similar to Claude Code, but extensible and transparent.


Features

Core Intelligence

• LangGraph Agent Loop: Deep agent architecture with planning, subagents, filesystem access, and tool-calling — all orchestrated through a shared async event loop
• Multi-Provider LLM Support: OpenAI, Anthropic, Ollama, Google Gemini — configure any provider via environment variables or the onboarding wizard
• Autonomous Learning System (Hermes): Periodically reviews tool usage patterns, extracts lessons, and autonomously creates reusable skills — the agent improves itself over time without user intervention
• Memory System: Two-tier persistent memory — auto-maintained markdown files (USER.md/MEMORY.md) plus a LangGraph key/value store (remember/recall) for cross-session facts
• Steering Instructions: Persistent user-defined directives injected into every model call — set once, always respected
• Inline Verification Loop: After each task, an out-of-band LLM call grades the output against a rubric — on a failing verdict, the agent is automatically re-driven with feedback (up to 3 retries). Fail-open by design
• Prompt-Template Hill Climbing: When reviews repeatedly flag the same class of misunderstanding, the system proposes a targeted rewrite of the relevant .jinja template, A/B tests it against the current version using verifier pass/fail as the quality signal,