Greptile uses GPT-6 Astra to test application behavior

OpenAI’s October 26 account describes testing proposed code changes by running the application. Greptile CEO Daksh Gupta gives the example of starting a development server, resizing its browser to an iPhone-sized view and using the model to spot a disappearing button. This is the companies’ account of the capability, rather than an independent test.

Greptile’s TREX documentation explains its testing workflow: identify affected user journeys, start the app in a configured, isolated test environment, then click, type and inspect the screen. It reports what was tested and what broke, with recordings. The organization enables TREX and configures an environment; changes without testable behavior can still be skipped.

The environment guide separates preparation from testing: dependencies are built into a saved snapshot from which reviews start. Instructions tell the agent which command starts the app, which port to open and how to sign in. Saving edited build settings does not build them; the user starts the build and checks its history. The recommended simulation needs a successful build matching those settings and runs without posting to pull requests.

The results guide distinguishes bugs from obstacles that prevented testing, such as a missing variable or a service failing to start. If no user journey was tested, that is not a passing result: the guide directs readers to inspect the blockers. Individual checks distinguish success, failure, blockage and work not attempted. Selecting a check in the flow view jumps to its moment in the recording.