danolongo/elefante ? reverse-engineered prompt

Build me a small coding agent in Rust that can inspect a project, run commands safely in an isolated sandbox, and remember useful project context between sessions.

I want it to feel like a practical helper for working inside a repository, so it can read files, suggest changes, run tests, and keep a memory of what it learned without mixing that memory into the project itself. It should be able to run in a headless mode from the terminal, and also have an interactive terminal UI that asks for approval before anything risky happens. If a sandbox is missing or cannot be started, it should fail closed and explain why instead of running commands unsafely.

Use a provider neutral design so the model backend can be swapped out easily, and make sure secrets never get exposed to the model or stored in memory. If you need current docs for any APIs or sandbox behavior, look them up online first.