ML-powered Python code change suggestion using the Coeditor transformer model.
Features
The extension provides the command Coeditor: Suggest edit below this line, which uses the cursor's line position to determine which Python function to perform the edit.
The extension is designed to work with Git projects and the model will condition its prediction on all the changes you made since the latest commit (plus any staged changes).
Inside the target function, only the changes you made before the cursor line (inclusive) are visible to the model, and it will make suggestions to the lines below the cursor.
You can also access the command from the editor context menu by right-clicking (available in .py files only).
Installing the Model
First, install the Coeditor model on your machine locally.
Then, start the suggestion service by running the command below inside the model's directory:
pipenv run python scripts/coeditor/start_server.py
The service will start on port 5042. You can specify a different port and GPU by modifying the script.
Note that currently, the server needs to be on the same machine as your source code in order for the server to query the git history.
Extension Settings
Server.Url: The URL of the suggestion service. Default to http://localhost:5042.
Usage Tips
If the model doesn't suggest the desired edit, try making some initial changes and moving the cursor to a lower position.
Any changes you made below the cursor will not be visible to the model. However, you can stage those changes first and the model will treat them as part of the last commit and make edits on top of those changes. You can selectively stage a range of changes easily using the VSCode UI.
This plugin only modifies a Python function that already exists in the last commit. It identifies a function using its full name (i.e., module_name.class_name.function_name). If you moved the function or changed its parent class's name, it would be treated as a new function and the model won't be able to suggest edits for it correctly (unless you stage thoese changes first).
If the extension is behaving unexpectedly, you can check what the model sees by looking at the server log files under <server_path>/coeditor_logs.
TODO: add a command to open the log files.
Known Issues
Currently, the service directly reads and writes to the files on disk, so the command Coeditor: Suggest edit from this line
will first save the file before calling the service.
When suggesting edits with drop_comments=True, all comments and doc strings are removed before feeding the code to the model. Applying the suggested chagnes back onto the orginal code with comments can be tricky and may not always work. We recommend using drop_comments=False for now.
Release Notes
0.2.3
Staged changes are now treated as part of the last commit.
0.2.2
The extension now tracks which suggestion has already been applied so the user can directly click another suggestion without having to undo the previously applied one.
