terminaltui
Star37

The font that spelled TERHINALTUI

3 min readascii, typography, bugs

                       _____   ___   ___   __  __    ___    ___   _   _   ___
                      |_   _| | __| | _ \ |  \/  |  / __|  / _ \ | \ | | | __|
                        | |   | _|  |   / | |\/| | | (__  | |_| || .\| | | _|
                        |_|   |___| |_\_\ |_|  |_|  \___|  \___/ |_|\_|  |_|

                         The Terminal Renaissance — June 15-16, Portland OR

────────────────────────────────────────────────────────────────────────────────────────────────────

  ❯ > Schedule [1]
     * Speakers [2]
     # Venue [3]
     $ Sponsors [4]
     ~ Home [5]

────────────────────────────────────────────────────────────────────────────────────────────────────
Real output, captured by booting the demo inside terminaltui's own headless PTY emulator.

terminaltui ships fourteen ASCII art fonts for page banners. One of them, Block, is the compact one — the face you get when the window is not wide enough for anything grander. It is four rows tall.

At four rows, M, V and W were all drawn the same way: a horizontal bar between two full-height stems, differing only in how high the bar sat.

That is not a cue anyone can read. Here is a real capture of a page set in it:

TRY: TERHINALTUI.DEU

That is supposed to say TRY: TERMINALTUI.DEV. The M rendered as an H. The V rendered as a U. From the one font compact enough to set a URL at mid-range window sizes, on the line telling people where to find the project.

WATCHING came out as HATCHING, which is at least funnier.

Why it survived

The failure mode is specific and it is worth naming, because it generalizes.

Banner text is display type. Nobody proofreads a headline — you look at it, recognize its shape, and move on. Body text gets read letter by letter, so a broken glyph there is caught in minutes. A broken glyph in a 4-row ASCII banner just makes the word look slightly wrong in a way you attribute to the medium. It is ASCII art. It is supposed to be crunchy.

And the tests were green, because the tests asserted that the banner rendered, that it was the right number of rows, and that it was the right number of columns. All true. The letters were the wrong letters.

The fix

All three are redrawn so the interior stroke changes column as it changes row, which is the only distinguishing signal available at this size:

  • M gains a descending V between the stems.
  • V narrows to a one-column point.
  • W gains a centre stem and two feet.

Each is one column wider than before, because centring a middle stem needs an odd ink width. That is a real cost — it means the compact face is now slightly less compact — and it is the correct trade, because a face that fits and lies is worth less than one that fits marginally worse and does not.

The test that should have existed

The interesting part is not the glyphs. It is that the test suite grew a new kind of assertion:

For every bundled face, render every letter, and fail if any two letters produce identical output.

That check does not need to know what an M should look like. It only needs to know that an M must not look like an H. It is cheap, it is total, and it catches the entire class of bug rather than the three instances we happened to find.

Most of the useful tests I have written lately have that shape. Not “is this output correct” — which requires knowing the answer in advance for every input — but “is this output self-consistent”, which the machine can check exhaustively without being told anything.