๐งช Having GPT-4 Iterate on Unit Tests like a Human
William Zeng - October 21th, 2023
Hi everyone, my name is William and Iโm one of the founders of Sweep.
Sweep is an AI junior developer that writes and fixes code by mirroring how a developer works.
1. Read the task description and codebase.
ClonedRepo is our wrapper around the Git API that makes it easy to clone and interact with a repo. We don't have any tests for this class, so we asked Sweep to write them.
Here Sweep starts by reading the original GitHub issue: โSweep: Write unit tests for ClonedRepoโ. https://github.com/sweepai/sweep/issues/2377 (opens in a new tab)
Sweep searches over the codebase with our in-house code search engine, ranking this symbol and file first: ClonedRepo:sweepai/utils/github_utils.py
.
This file sweepai/utils/github_utils.py (opens in a new tab) is ~370 lines long, but because we know the symbol ClonedRepo
, we extracted the relevant code (~250 lines) without the other functions and classes.
import git
# more imports
...
class ClonedRepo:
repo_full_name: str
installation_id: str
branch: str | None = None
token: str | None = None
@cached_property
def cache_dir(self):
# logic to create a cached directory
# other ClonedRepo methods
def get_file_contents(self, file_path, ref=None):
local_path = os.path.join(self.cache_dir, file_path)
if os.path.exists(local_path):
with open(local_path, "r", encoding="utf-8", errors="replace") as f:
contents = f.read()
return contents
else:
raise FileNotFoundError(f"{local_path} does not exist.")
# other ClonedRepo methods
We read this to identify the necessary tests.
2. Write the tests.
Once Sweep has the reference implementation, Sweep generates the corresponding test as commits in a GitHub PR (opens in a new tab):
def get_file_contents(self, file_path, ref=None):
local_path = os.path.join(self.cache_dir, file_path)
if os.path.exists(local_path):
with open(local_path, "r", encoding="utf-8", errors="replace") as f:
contents = f.read()
return contents
else:
raise FileNotFoundError(f"{local_path} does not exist.")
We have Sweep generated mocks for os.path.join
and open
.
This code looks great!
@patch("os.path.join")
@patch("open")
def test_get_file_contents(self, mock_open, mock_join):
mock_join.return_value = "/tmp/cache/repos/sweepai/sweep/main/file1"
mock_open.return_value.__enter__.return_value.read.return_value = "file content"
content = self.cloned_repo.get_file_contents("file1")
self.assertEqual(content, "file content")
We generated mocks for os.path.join
and open
, which should return the correct path and file contents.
Ok we're done here right? Can we just write these tests and leave the rest to the developer?
3. Run the tests.
Most other AI tools stop here, but itโs not enough.
If you just committed these tests it would be great, but youโd end up with a frustrating bug. Here it is:
File "/usr/lib/python3.10/unittest/mock.py", line 1616, in _get_target
raise TypeError(
TypeError: Need a valid target to patch. You supplied: 'open'
Did we really save time for the developer here? Itโs frustrating that most other tools donโt fix these issues.
Unlike every other tool, Sweep actually runs these tests.
Sweep ran the code, found the issue, and identified the solution:
โChange the target of the patch in the 'test_get_file_contents' method from 'open' to 'builtins.open'. This will correctly patch the built-in 'open' function during the test.โ
Sweep added this commit (opens in a new tab):
@patch("os.path.join")
- @patch("open")
+ @patch("builtins.open")
def test_get_file_contents(self, mock_open, mock_join):
We ran the tests again, and finally we have unit tests that actually work! ๐
Visualizing Sweep's process
Sweep updates you with a flowchart showing you the code it's written and the tests it's run.
Sweep starts off by writing the original github_utils_tests.py
file.
Then Sweep runs the tests using python sweepai/utils/github_utils_test.py
. (this command is configurable)
Sweep found the error, and suggests a new modification to the tests.
Finally, we used the test feedback to fix the code and run the tests again. If there was a bug in the original github_utils.py
, we would have fixed it here.
Becoming an AI junior developer
The benefits of this approach seem small at first: โOk, now I donโt need to fix the testsโ. But if you just wanted to write unit tests, you might as well use ChatGPT.
With this approach, Sweep can not only write the test, but:
- find bugs in the original code
- fix the original code
- run the tests again to confirm the bugfix
One step closer to becoming an AI developer.