T3 Code: debugging GPU load with Fable and Codex

This debugging walkthrough follows the speaker's investigation of excessive resource use in T3 Code, an open source alternative to the Codex app. He uses its browser interface to work with coding tools on other computers. Although the interface remained responsive, he reports that the browser's GPU process consumed up to 50% CPU at higher display resolutions.

An initial Codex rewrite targeted network handling and React updates without meaningfully improving his experience. Fable also pursued explanations he found unhelpful. The useful change in approach was to ask the coding assistant to build experiments: console helpers that could disable CSS effects and reset the page for comparison.

Those tests pointed to continuous opacity animations, including pulsing terminal icons in the sidebar. The speaker ultimately traced the problem to their interaction with backdrop blur and a page-wide noise layer, especially on his high-resolution, 120 Hz setup. Removing the noise required further color adjustments, which he found particularly noticeable on HDR displays.

The walkthrough also explains why he isolated T3 Code in a separate Chrome window. Other idle Claude pages later obscured the improvement in Helium. These measurements describe his setup, rather than universal browser performance. His practical lesson is to use an AI agent to search code and build diagnostic tools while testing its explanations against observed behavior.