Native transcript rendering
Answers now render as documents rather than as markdown source. Tables become
real tables, fenced code gets a container, inline HTML is honoured, and LaTeX
is typeset. The transcript styles each block as it finishes streaming, so what
you read while the answer arrives is what you are left with when it ends.
Nothing new to install, and the previous behaviour is one environment variable
away.
Notable
- Tables render with per-column alignment and a weighted header row. A table
anywhere in an answer no longer sends the whole message to raw text, and a
row with blank or missing cells still draws a full grid. - Math typesets from
$...$,$$...$$,\(...\),\[...\]and the KaTeX
block environments. Currency, shell variables and$inside link
destinations are left alone. - An equation that cannot be typeset stays visible as its own source. It is
never blank, and never drawn with a character silently missing — the pinned
engine drops characters it has no atom for without raising an error, so the
renderer refuses those before drawing. - Fenced code renders as code wherever it sits: at the top level, under a
bullet, inside a quote, or nested two items deep. - Copy carries the LaTeX, not the picture of it. Cmd-C, drag and a Services
provider all write the same bytes; a fifty-equation answer went from 12.3 MB
of rasterised images to 36 KB of text. - A selection survives a streaming tick. Only the stretch that actually changed
is reported as rewritten, so text above the open block keeps its indices. TURBO_FIELDFARE_DISABLE_MATH=1restores raw LaTeX, and
TURBO_FIELDFARE_PROGRESSIVE_RENDER=0restores raw streaming.