nexustech101/ollama-harness ? reverse-engineered prompt

Build me a small Python coding agent that I can run from the terminal. I want it to chat with a model, use local project files, search, shell, and git tools when needed, and stream its responses as it works so it feels live. It should have a simple command line REPL for chatting in the current folder, plus a way to run one prompt from the command line like fixing or editing a file.

Also add a server mode that exposes the same agent through an OpenAI compatible HTTP API, including streaming responses and a models endpoint. I want it to work with Ollama by default, but also support OpenAI, OpenRouter, and other OpenAI compatible endpoints through a provider config file and environment variables. If possible, make the provider setup easy from the command line too.

Please keep it clean and small, with sensible defaults, and make sure the terminal version and API version share the same agent loop.