This work was envisioned several years ago, discussed endlessly between different people in the project, gathering requirements and trying to come up with a useful design. In a very real sense, this really started with @soph's work over a decade ago, when she led the work to overhaul the design vocabulary of Rust diagnostics https://blog.rust-lang.org/2016/08/10/Shape-of-errors-to-come/. Her original work still looms and underpins everything that has come since, and the flexibility of that design allowed for separating the semantic content of errors from their rendering, which meant that changing how things *look* doesn't require any changes to `rustc`. The work we've already done in evolving the content of diagnostics, like shown in https://kobzol.github.io/rust/rustc/2025/05/16/evolution-of-rustc-errors.html and that you could see for yourself by running https://git.donsz.nl/jana/errorversion, already does the heavy lifting.
I recall starting conversations about having an output suitable using assistive technologies like screen-readers on almost every RustConf and RustWeek I've attended, specially spending time with @epage and @muscraft hunched over laptops on cardboard tables or arguing over tikka masala and beers on what the best design could be. There's plenty of prior art, but one example is from @zkat's Miette, another diagnostic renderer for Rust: https://toot.cat/@zkat/115584770971710617. I *believe* they are the ones that came up with the `NO_GRAPHICS=1` environment flag, mimicking `NO_COLOR=1` https://no-color.org/.