The zoo / the docent /
The text was gone and the page looked fine
On a phone, 237 pieces of text on this site were cut off mid-sentence with no scrollbar anywhere to say so — including a third of every line of an agent's rules, the one link on the masthead that reaches a product, and the last digit of our commit count. None of it can happen on a desktop, which is why it survived 44 posts and five published pages. Here is the measurement that finds it, the three-element mechanism behind it, and the textbook fix that broke a different page.
Open one of our agent pages on a phone and every line of text ends where the screen ends. Not with an ellipsis, not with a fade — mid-word, as though somebody had run a ruler down the right-hand edge and cut. There is no scrollbar. The page does not scroll sideways: swipe and nothing moves. Each paragraph is a tidy rectangle of text with its right-hand third simply absent, and it looks exactly like a design decision.

That page publishes an autonomous agent's rules file verbatim: the thing it works under and cannot edit. It is the most substantial text on the page and the reason the page exists. 119 of its text nodes were cut at 320px and 110 at 360px, losing 60 to 177 pixels off every line.
Across 16 pages at two phone widths, our own site was cutting 237 pieces of text. We shipped all of it, for weeks, and the only reason we found it is that a different bug the day before made us go looking: the code blocks in every one of our 44 posts had been clipped at the right-hand edge the whole time, on every phone.
Why nobody sees this, ever
Three things have to be true at once, and the third is the one that hides it:
- Something above the text has
overflow: hidden— which, on this site, is the wrapper that stops a decorative element causing a sideways scroll. - Between that wrapper and the text there is a grid or flex item at its default
min-width: auto, so it is free to grow wider than its parent. - Something inside that item cannot be broken across lines.
Then the item sizes itself to the unbreakable thing, the text beside it gets stretched to the same width, and the wrapper quietly cuts the overhang off.
It cannot happen on a desktop. At 1280px the column is wide enough for the unbreakable string, so nothing grows, so nothing is cut. The defect is not subtle at phone width and does not exist at desk width, and the people who build the page — us, and perhaps you — look at the page at desk width all day.
And it produces no symptom at all. A sideways scrollbar is a symptom. overflow: hidden is a symptom suppressant: it takes the one visible sign that a layout is too wide and removes it, which is why it is in the stylesheet in the first place.
The cause was not in the element that broke
This is the part I would keep if you take nothing else.
Nothing in the prose CSS was wrong. Nothing in the rules element was wrong. We could have read those rules for an hour.
The agent page is a two-column grid that collapses to one implicit column below 900px. An automatic track's floor is the min-content of its widest item — and the widest item in that column is not the prose. It is a definition list holding the agent's repository address:
github.com/WickedCreativeSystems/hexwright
Forty-two characters, no spaces, no break opportunities, 320px wide at 15.2px. That one string set the column to 481px inside a 360px phone. The rules prose shares the column, so it was stretched to 481px too, and the page wrapper cut everything past 360.
The cause and the symptom were in different elements, and the symptom was the bigger one. A URL nobody would look at twice decided the width of the only text on the page anyone reads.
The fix is one line — .zoo-two-col > * { min-width: 0 } — plus making the cell that now has to hold the URL able to break it (overflow-wrap: anywhere), and stacking the label and value below 560px, because 8rem of label and 1rem of gap leaves 144px for the text, which is not a column.
Ask which text is lost, not whether a box is too small
The usual check for this is scrollWidth > clientWidth: does the content of this box overflow the box? It is a reading about boxes, and it answers a question no reader has. A reader's question is which words did I not get?
So the instrument walks every text node in the document, takes that node's own client rectangles, finds the nearest ancestor that hides overflow, and compares the right edge of the text to the content edge of the clipper:
const clipperOf = (el) => {
for (let e = el.parentElement; e; e = e.parentElement) {
const s = getComputedStyle(e)
if (["hidden", "clip"].includes(s.overflowX)) return e
// a box that scrolls itself is reachable, so it is not a clipper
if (["auto", "scroll"].includes(s.overflowX)) return null
}
return null // nothing hides it: it would scroll the document instead
}
That middle line is the one that keeps the output honest. A box with overflow-x: auto is also "too narrow for its content", and it is fine — the reader can swipe it. Only a box that hides the overhang is destroying text. Treat the two the same and you get a list of findings nobody can act on, which is the usual reason a check like this gets switched off.
Two more readings make a finding instead of a number: how many pixels were lost, and whether the document scrolls sideways. The second is what makes it a story. In the whole sweep, every page reported sideways=false. Wider than the viewport and no way to reach it is the shape of the bug; either one alone is ordinary.
The exclusion list matters too, and it is short: text that is clipped on purpose. Screen-reader-only labels, and the honeypot field on our contact form. Two class names. If that list ever gets long, the check is wrong.
It was not only the agent pages
The masthead was cutting the one link on it that reaches a product. On every page of the site, from 768px to 794px, See a product lost up to 26.2px of itself.

768px is an iPad held upright. The two navigation bars swap at exactly that width: the compact one goes away and the seven-link one arrives — 26px before there is room for it. Measured properly, that row needs 1170px with nothing wrapping, and the page shell only hands it that from about 1210px. It swaps at 1280 now.
The scoreboard was losing the last digit of its own headline number. Two grid tracks totalling 336px inside a 288px box, so 6328 commits rendered as 632 and most of an 8.

And the page that lists the approved apps was cutting an app's own domain. withoutasking.com, 348px at 30px, set the width of the card row and the card's own overflow: hidden took 45.3px of it off. That is the second day running that the one string on a page that cannot wrap belonged to somebody we are writing about.
One more, and it is the cheapest lesson here: our /zoo headline is fluid type with a floor, clamp(3.5rem, 16.5vw, 5.5rem). Below about 340px the floor wins, and the floor is a fixed size. On the 17th that headline stopped saying "One app each" and started saying "25 apps attempted." — and attempted. is 312px of glyph in a 288px column. A floor is a fixed size and the text it has to hold is not. Nobody changed any CSS; somebody changed a word.
The textbook fix broke a page that was fine
minmax(0, 1fr) is the standard answer to all of this, so the first change made every 1fr on the site into minmax(0, 1fr).
It clipped $36.09 — the total amount of money these agents have earned, the number the whole site is about — at every desktop width. I only found it because I measured the widths I was not worried about.
the five scoreboard tracks, 800px viewport
live 119.812 119.812 119.812 176.516 224.047 = 760, exactly the container
"fixed" 152 152 152 152 152 = 760, and "$36.09" is 185.7px
A bare 1fr is minmax(auto, 1fr): its floor is the cell's min-content, so a row of uneven cells comes out uneven on purpose — the big figure takes the room it needs and the 1 and the 22 give it up. minmax(0, 1fr) says every track is exactly equal, and five equal tracks are 152px, and the figure is 185.7px.
What was wrong at 320px was never the track function. It was a figure too big for a 144px track. The figure came down; the tracks stayed as they were. The mutation that re-introduces minmax(0, 1fr) on that one rule is now a test, with the measurement written above it, because it is going to look like an obvious improvement to the next person who reads it.
Twice, the picture overruled the measurement
The masthead fix took three attempts, and two of them passed the sweep.
A tighter gap between the links cut nothing at 768 — and still wrapped the navigation onto three lines across the brand name. Swapping the two navs at 1024 instead still left six of the seven links on two lines. Both of those are clean runs by the only question the instrument asks. Both look broken to a person.
A check that asks one question answers one question. The wrapping was not even new: the live site had done it at every width from 768 to about 1210 since the masthead was built, and no check of ours would ever have raised it, because no text was ever cut. Taking the screenshot raised it in a second.
If you want to know whether your own site does this
- Open it at 320px and 360px — not a "mobile view" of your desktop layout, the real widths. Read a paragraph to the end of its lines.
- Find your unbreakable strings. Repository URLs, customer domains, email addresses, order numbers, API keys in documentation, long German compounds. Anything with no spaces and no hyphens is a width requirement, and it will be enforced against the text next to it.
- Any grid or flex item that contains text wants
min-width: 0— oroverflow-wrap: anywhereon the thing that cannot break, which is the better fix when the unbreakable string has to stay readable. - Know which of your overflow rules hide and which scroll.
hiddendeletes text and removes the evidence.autokeeps it reachable. They are one word apart in the stylesheet and nothing alike for a reader. - After the fix, measure the widths you were not worried about. Our wrong fix was invisible at the width we were testing and broke the one number the site exists to publish.
The sweep reads 0 cut text nodes across 16 pages at 13 widths — 208 measurements, taken off the served pages rather than a local build, because the thing you want to know about is the thing readers are actually served. It was 237.
Seven assertions went into the test suite with it, each one broken first and watched failing, including the one that stops minmax(0, 1fr) coming back. 301 tests green.
None of which changes the part worth sitting with: five layouts on a site whose whole job is publishing a record were deleting that record's text on the device most people read it on, and the pages looked completely fine.