Analysis
A button disappears at an iPhone-sized browser viewport. An October 26 OpenAI customer story says Greptile starts the application for a pull request, changes the viewport and uses GPT-6 Astra to examine the screen. The story is a company account of that workflow. It is not an independent evaluation.
The viewport is the setting that decided whether the button was on the screen.
A useful finding names the tested revision, the viewport, the steps and the visible failure, so another reviewer can reproduce it. Browser exploration can reveal a broken journey. It cannot by itself prove security, accessibility or coverage of every important path. Keep those acceptance checks explicit.
The same account says Greptile supplies codebase documentation to Codex through a Model Context Protocol server. Greptile says it tracks reversions of AI-generated code and sees fewer when agents receive that context.
Not established: a numerical reduction, a controlled comparison, or a security certification. Fewer reversions, on this evidence, is a testing question. It is not a quantified promise.
A screen the model can describe is not a trust boundary.
Three questions for the person who runs the system. Does each finding name the revision, the viewport, the steps and the visible failure? Which security, accessibility and path checks stay explicit, outside the browser pass? Who owns the documentation the agent receives, and where is the off switch?
Source: Greptile takes bug detection beyond code review with GPT-6 Astra · 2026-10-26
Why it matters
A model can show you a broken screen and still leave security, accessibility and path coverage unwritten.
