AutoPatch API
Finds code that breaks when a third-party API changes, and proposes safe patches.
- Date
- Aug 2026
- My role
- Built it with an AI coding assistant (commits are co-authored)
- Visibility
- Public
Details
Example case: Stripe API version 2020-03 to 2020-08.
It diffs API specs with oasdiff and guesses renames.
A regex scanner is the baseline. An AST scanner raised precision from 0.26 to 0.41 at the same recall (13 of 13) on 32 hand-labelled cases.
Tool-using LLM agents judge impact and generate patches. A validator checks patch safety.
It verifies before and after against a Prism mock.
Decisions and why
- Terminal submit tools instead of free-form JSON answers.
- Gives agents a clear, checkable way to finish.
- Measure against a regex baseline.
- Without a baseline, there is no way to know if the AST scanner helps.