Local CodeLlama Engineering Workbench

A privacy-oriented coding assistant that routes source code into five explicit engineering workflows and runs entirely against a local Ollama endpoint.

Distinct use case

This repository is not a general chat application and does not use hosted RAG. It is a local code-review workbench for:

• explanation
• debugging
• test generation
• behavior-preserving refactoring
• security review

Each workflow receives a different system instruction and produces a predictable four-part response.

Architecture

flowchart LR
    A["Source code"] --> B["Task router"]
    B --> C["Structured prompt"]
    C --> D["Local Ollama / CodeLlama"]
    D --> E["Findings, patch and risks"]

Loading

Run

ollama pull codellama
ollama serve
git clone https://github.com/sandeep848/codellama-dynamic-prompt-handler.git
cd codellama-dynamic-prompt-handler
pip install -r Codellama/requirements.txt
streamlit run Codellama/app.py

No cloud API key is required and pasted code stays on the machine hosting Ollama.

Test

pytest

Key distinction

The routing layer is deterministic and inspectable. The application shows the exact prompt sent to the model and explicitly states that generated feedback has not been executed.