Fix
Hand it a stack trace, a GitHub issue or an error Sentry saw. It finds the cause, proves it with a failing test, and proposes a patch you approve, verified by the tests.
- Nothing is patched until a test reproduces the bug; a flaky reproduction stops the run.
- A person approves the plan before any code changes; a risky one needs two.
- Tests run in a sandbox with no internet access; only installing dependencies reaches the package registries.
- A patch counts only when the reproduction passes and no test that passed before fails.
- Changes to tests, skipped tests or a faked report are caught and refused.
- Changes in plain words go the same way: planned by layer, approved, patched with new tests.
How a fix runs, stage by stage
- Intake (done)
- Investigate (done)
- Reproduce (done)
- Root cause (done)
- Plan (done)
- Approve (done)
- Patch (done)
- Verify (current)
- PR (to come)
1
Intake
The trace, issue or Sentry error is read into frames in your code.
2
Investigate
An agent reads the code and its history to find the cause.
3
Reproduce
It writes a test that fails for that reason, and runs it three times.
4
Root cause
What else the cause reaches, from the code's graph.
5
Plan
The files and steps of the fix, with its risk.
6
Approve
People with write access approve it, ask for changes, or reject it.
7
Patch
The plan is carried out in the sandbox, file by file.
8
Verify
The whole suite runs again, and the patch is checked for gaming.
Opening the pull request on your repository, with the patch, its test and the evidence, is coming soon.