adam-s/goldseam

goldseam is a testing tool for Cypress, a popular browser testing framework, that helps fix broken tests when a website's UI changes. When a selector used to find an element on a page breaks, goldseam captures a redacted snapshot of the page's structure, sends it to an AI model of your choice, and asks the model to propose a minimal fix. That fix is checked through a multi-step verification process, and only if it passes does it become a small, reviewable code change you can inspect before accepting. A key idea behind the project is that you bring your own AI model rather than relying on a vendor's cloud service. It works by default with the Claude Code command line tool many developers already have installed, but also supports a local Ollama model, any OpenAI-compatible API endpoint, or any custom command line program. This means test data and application code never have to leave your own machine or chosen service. goldseam also has a second mode called authoring, which lets you write test steps in plain English, such as adding an item to a cart and checking the cart count. These plain-English steps are translated once into real Cypress commands by the model, and the result is cached as a file that gets committed to the repository. Because the translation is cached, the test suite can run in continuous integration afterward without making any further calls to the AI model. The project is careful about honesty in its results. Every proposed fix only changes selector strings, never test logic or assertions, and if the tool cannot confidently repair something, it reports a give up rather than pretending success. Tests that already pass are left completely untouched, and the tool behaves the same whether or not it's installed, until a failure actually happens. The repository is organized as a set of packages, including the main goldseam plugin and command line tool, a standalone accessibility snapshot utility, a demo shop application used for testing, and a folder of real-world tests used to prove the tool works against actual applications.