patchllm 0.2.2

Lightweight tool to manage contexts and update code with LLMs

About

PatchLLM is a command-line tool that lets you flexibly build LLM context from your codebase using glob patterns, URLs, and keyword searches. It then automatically applies file edits directly from the LLM's response.

Usage

PatchLLM is designed to be used directly from your terminal. The core workflow is to define a scope of files, provide a task, and choose an action (like patching files directly).

1. Initialize a Scope

The easiest way to get started is to run the interactive initializer. This will create a scopes.py file for you, which holds your saved scopes.

patchllm --init

This will guide you through creating your first scope, including setting a base path and file patterns. You can add multiple scopes to this file for different projects or tasks.

A generated scopes.py might look like this:

# scopes.py
scopes = {
    "default": {
        "path": ".",
        "include_patterns": ["**/*.py"],
        "exclude_patterns": ["**/tests/*", "venv/*"],
        "urls": ["https://docs.python.org/3/library/argparse.html"]
    },
    "docs": {
        "path": "./docs",
        "include_patterns": ["**/*.md"],
    }
}

2. Run a Task

Use the patchllm command with a scope, a task, and an action flag like --patch (-p).

# Apply a change using the 'default' scope and the --patch action
patchllm -s default -t "Add type hints to the main function in main.py" -p

The tool will then:

1. Build a context from the files and URLs matching your default scope.
2. Send the context and your task to the configured LLM.
3. Parse the response and automatically write the changes to the relevant files.

All Commands & Options

Core Patching Flow

• -s, --scope <name>: Name of the scope to use from your scopes.py file.
• -t, --task "<instruction>": The task instruction for the LLM.
• -p, --patch: Query the LLM and directly apply the file updates from the response. This is the main action flag.

Scope Management

• -i, --init: Create a new scope interactively.
• -sl, --list-scopes: List all available scopes from your scopes.py file.
• -ss, --show-scope <name>: Display the settings for a specific scope.

I/O & Context Management

• -co, --context-out [filename]: Export the generated context to a file (defaults to context.md) instead of running a task.
• -ci, --context-in <filename>: Use a previously saved context file as input for a task.
• -tf, --to-file [filename]: Send the LLM response to a file (defaults to response.md) instead of patching directly.
• -tc, --to-clipboard: Copy the LLM response to the clipboard.
• -ff, --from-file <filename>: Apply patches from a local file instead of an LLM response.
• -fc, --from-clipboard: Apply patches directly from your clipboard content.

General Options

• --model <model_name>: Specify a different model (e.g., gpt-4o). Defaults to gemini/gemini-1.5-flash.
• --voice: Enable voice recognition to provide the task instruction.

Setup

PatchLLM uses LiteLLM under the hood. Please refer to their documentation for setting up API keys (e.g., OPENAI_API_KEY, GEMINI_API_KEY) in a .env file and for a full list of available models.

To use the voice feature (--voice), you will need to install extra dependencies:

pip install "speechrecognition>=3.10" "pyttsx3>=2.90"
# Note: speechrecognition may require PyAudio, which might have system-level dependencies.

License

This project is licensed under the MIT License. See the LICENSE file for details.