Patch-Pal

Patch-Pal is a small Boot.dev project that wires Gemini tool-calling to a local sandbox. It exposes a handful of file and process utilities as tools and lets the model request them in a loop until it returns a final answer.

What’s included

• A chat runner that manages the Gemini request/response loop, tool calls, and conversation history.
• Tool functions for listing files, reading files (with truncation), writing files, and running Python files.
• A simple CLI entry point.
• Test scripts that exercise the tool functions directly.

Project layout

• main.py CLI entry point.
• chat_runner.py Gemini request/response loop.
• prompt.py system prompt used for the model.
• tools.py tool registry and thin dispatcher (legacy).
• functions/ LLM-callable functions and their schemas.
• config.py configuration (e.g., MAX_CHARS for file reads).
• calculator/ sample project the tools operate on.

Requirements

• Python 3.11+ (uses a local .venv in this repo)
• uv for running scripts
• A Gemini API key in .env:
  • GEMINI_API_KEY=...

env

cp .env.example .env

Run

uv run main.py "run tests.py" --verbose

Tool tests

uv run test_get_files_info.py
uv run test_get_file_content.py
uv run test_write_file.py
uv run test_run_python_file.py

Notes

• Tool functions always return strings and never raise errors to the model.
• File reads are truncated at MAX_CHARS (see config.py).
• The chat loop stops when the model returns a final response, or after a max iteration limit.