GitLocker: The Coding Marketplace

Description:

autobotml 0.0.16

autobot

An automated code refactoring tool powered by GPT-3. Like GitHub Copilot, for your existing
codebase.
Autobot takes an example change as input and generates patches for you to review by scanning your
codebase for similar code blocks and "applying" that change to the existing source code.

See more examples on
Twitter, or read the
blog post.
N.B. Autobot is a prototype and isn't recommended for use on large codebases. See: "Limitations".
Getting started
Autobot is available as autobot-ml on PyPI:
pip install autobot-ml

Autobot depends on the OpenAI API and, in particular, expects your OpenAI
organization ID and API key to be exposed as the OPENAI_ORGANIZATION and OPENAI_API_KEY
environment variables, respectively.
Autobot can also read from a .env file:
OPENAI_ORGANIZATION=${YOUR_OPENAI_ORGANIZATION}
OPENAI_API_KEY=${YOUR_OPENAI_API_KEY}

From there, you can run any of Autobot's built-in refactors (called "schematics"):
autobot run useless_object_inheritance /path/to/file.py

Example usage
TL;DR: Autobot is a command-line tool. To generate patches, use autobot run; to review the
generated patches, use autobot review.
Autobot is designed around a two-step workflow.
In the first step (autobot run {schematic} {files_to_analyze}), we point Autobot to (1) the
"schematic" that defines our desired change and (2) the files to which the change should be
applied.
In the second step (autobot review), we review the patches that Autobot generated and, for each
suggested change, either apply it to the codebase or reject the patch entirely.
Autobot ships with several schematics that you can use out-of-the-box:

assert_equals
convert_to_dataclass
numpy_builtin_aliases
print_statement
sorted_attributes
standard_library_generics
unnecessary_f_strings
use_generator
useless_object_inheritance

For example: to remove any usages of NumPy's deprecated np.int and associated aliases, we'd first
run autobot run numpy_builtin_aliases /path/to/file.py, followed by autobot review.
The schematic argument to autobot run can either reference a directory within schematics (like
numpy_builtin_aliases, above) or a path to a user-defined schematic directory on-disk.
Implementing a new refactor ("schematic")
Every refactor facilitated by Autobot requires a "schematic". Autobot ships with a few schematics
in the schematics directory, but it's intended to be used with user-provided schematics.
A schematic is a directory containing two files:

before.py: A code snippet demonstrating the "before" state of the refactor.
after.py: A code snippet demonstrating the "after" state of the refactor.

Each file is expected to consist of a brief top-level docstring describing the "before" or "after"
state, followed by a single function or class.
For example: in Python 3, class Foo(object) is equivalent to class Foo. To automatically remove
those useless object inheritances from our codebase, we'd create a useless_object_inheritance
directory, and add the following two files:
# before.py
"""...with object inheritance."""
class Foo(Bar, object):
def __init__(self, x: int) -> None:
self.x = x

# after.py
"""...without object inheritance."""
class Foo(Bar):
def __init__(self, x: int) -> None:
self.x = x

We'd then run autobot run ./useless_object_inheritance /path/to/file/or/directory to generate
patches, followed by autobot review to apply or reject the suggested changes.
Limitations

Running Autobot consumes OpenAI credits and thus could cost you money. Be careful!
By default, Autobot uses OpenAI's text-davinci-002 model, though autobot run accepts a
--model parameter, allowing you to select an alternative OpenAI model. Note, though, that
OpenAI's Codex models are currently in a private beta, so code-davinci-002 and friends may
error for you.
To speed up execution, Autobot calls out to the OpenAI API in parallel. If you haven't upgraded
to a paid account, you may hit rate-limit errors. You can pass --nthreads 1 to autobot run
to disable multi-threading. Running Autobot over large codebases is not recommended (yet).
Depending on the transform type, Autobot will attempt to generate a patch for every function or
every
class. Any function or class that's "too long" for GPT-3's maximum prompt size will be skipped.
Autobot isn't smart enough to handle nested functions (or nested classes), so nested functions
will likely be processed and appear twice.
Autobot only supports Python code for now. (Autobot relies on parsing the AST to extract relevant
code snippets, so additional languages require extending AST support.)

Roadmap

Multi-language support. Autobot only supports Python code for now. Extending to
multi-language support, at least with the current algorithm, will require supporting additional
AST parsers. The most likely outcome here will either be to leverage tree-sitter.
Supporting large codebases. What would it take to run Autobot over hundreds of thousands of
lines of code?

License
MIT