Hey everyone,
I got frustrated with AI coding assistants blindly applying edits, so I built Nova CLI—a terminal-based agent designed to handle long, tedious tasks with actual engineering rigor.
It supports cloud providers (OpenAI, Anthropic, Gemini), but it's fully built to run on local Ollama models for completely offline, private coding.
🧠 What makes it different (The 10 Pillars)
Instead of just writing code and stopping, it has a multi-step verification pipeline:
Sentinel (Adversarial Red Team): After the main agent finishes writing code, it automatically spawns a hidden subagent. The subagent's only job is to try and break the code that was just written before the AI tells you it's "done."
Seismograph (Regression-Risk): Before editing a file, it checks git log to see if that exact code block has been volatile. It calculates a 0-100 risk score and warns you if the code has been reverted multiple times.
Time-Machine (Git-Blame RAG): Uses ChromaDB for semantic code search, fuses the results with git blame so the AI knows why code was written.
Architect (Impact Analysis): Builds a real AST dependency graph of all callers of a function before making breaking changes.
Ground-Truth (Environment Drift): AST-parses your codebase for imports and warns if you're writing code for a package that isn't installed.
Echo-Loop (Runtime Verification): Runs the generated Python code in an isolated subprocess with a 10-second timeout to verify it actually executes.
Oracle (Confidence Scoring): Runs unittest blocks against generated code and calculates a true graduated 0-100% pass rate.
Hivemind (Shared Team Memory): Auto-commits and pushes team preferences (memory.md) to Git.
Show-Your-Work (Diff Annotations): Tags every git diff line [ADDED]/[REMOVED] and forces the model to explain them.
Reverse-Engineer (Binary Analysis): Uses Capstone to disassemble .exe and .so files into assembly.
It also has built-in local Whisper for audio transcription (runs entirely offline) and voice control (/mic).
💻 How to get it
I just compiled a standalone Windows .exe so you don't even need Python installed to try it.
GitHub: https://github.com/adomba102-png/advcli
Install: Go to the Releases page, download the zip, extract, and run nova-cli.exe. Type /provider to select Ollama!
⚠️ The Honest Disclaimer
This is an alpha-stage prototype built by one developer. It has bugs. The "sandbox" is a temporary directory with a timeout, not a true Docker container. But the architecture is real, the tools are real.
I'd love for you to break it and tell me what goes wrong. Check the code out and let me know what you think! You can find the short at : https://www.youtube.com/shorts/v93GpVewz58
Top comments (0)