Repository navigation
MAIN - #17280
Open
niteeshkanna-sh wants to merge 388 commits into
Open
MAIN#17280niteeshkanna-sh wants to merge 388 commits into
niteeshkanna-sh wants to merge 388 commits into
Conversation
|
Hi @niteeshkanna-sh! Thank you for your pull request and welcome to our community. Action RequiredIn order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you. ProcessIn order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA. Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with If you have received this in error or have any questions, please contact us at cla@meta.com. Thanks! |
Fourteen places showed an image and only some of them could be changed without a deploy. Two separate causes. **SectionArt never asked the panel.** It looked in public/photos and nowhere else, so the six cards under "What we hire", the three "How it works" steps and the areas-we-serve panel all ignored anything uploaded. That is worse than having no slot: someone uploads a photograph of their own fleet, the panel says it saved, and the drawing stays. It asks useSiteImage now -- uploaded first, then a file in the repository, then the drawing -- which also makes the special case AreasServed carried for `coast` unnecessary. **Three more were hardcoded paths**: the home page background, the two photographs beside "Why hire from us", and the open road band. The first picture a visitor sees was the one picture the owner could not change. Website content gains thirteen slots: Home page background, the two Why-hire photos, Open road band, the six cards named after the page each links to, and the three steps. Nineteen images on the site, every one of them uploadable. Slot names may now contain digits. They could not: the sanitiser stripped them and the reader's pattern rejected them, so step-1, step-2 and step-3 would all have been written as "step-" and then served as a 404. The open road band stays a CSS background rather than becoming an <img>. If the file is missing the band falls back to the navy beneath it and still looks deliberate, where an <img> would leave a broken-image icon on a live page. Verified by seeding all thirteen new slots: every one renders, no drawing is left on the home page, the six cards carry their uploads on /services too, and nothing overflows at 390px. The seeds are not committed. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Every picture on the site is uploadable from the panel
**The grid was ragged because of a control nobody chose the width of.** Each
slot carried a bare <input type="file">, which prints "Choose File — No file
chosen" at whatever width the browser feels like. With nineteen of them on one
page no two rows agreed where anything sat.
The input is still there and still the thing that opens the file chooser; it is
just no longer the thing anyone looks at. Every card is now the same width,
its buttons pinned to the bottom so they line up across a row whatever the
label above them did.
**Every image can be framed before it is uploaded, and re-framed afterwards.**
Drag to move, a slider and a pair of buttons to zoom. Whatever is inside the
frame is what the site shows -- rather than the browser cropping a photograph
of some other proportion and nobody knowing which part survived until the page
loads.
The frame is the shape the site actually lays that slot out in, and each is
saved at a size to match:
logo, snake 1:1 512x512
the page banners 16:5 1600x500
home hero, open road 16:9 1600x900
the why-us photos 4:3 1200x900
cards and steps 16:10 1200x750
One crop for all of them would be wrong in both directions: a logo squeezed
into a letterbox loses its top and bottom, a banner squared off loses its
sides.
A slot with an image gets Replace, Edit and Remove; an empty one gets Choose
image. Edit re-frames what is already stored, which works because the panel
serves it from this same origin -- a cross-origin source would taint the canvas
and toBlob would throw.
The framed result goes into the hidden input through a DataTransfer, so this
stays an ordinary form post rather than a fetch with its own error handling.
Remove sets the action on that same form: a form cannot contain another, and a
second one beside it would be a second cell in the grid.
Also fixes a rule of mine from the mobile pass. `input[type="file"] { width:
100% }` outranked the visually-hidden class and stretched every hidden input to
the full width, pushing a 390px page out to 447. It now excludes the one input
that is not a visible control.
Verified: 28 cards, all one width, no native file control drawn, three to a row
on a desktop and one on a phone, nothing over 320px. Choosing a file opens a
frame measuring 1.00 for the logo and 3.20 for a banner; the zoom buttons move
100 to 140; saving puts a webp in the form and posts brand-upload; Edit reopens
the stored image and posts a new one; Remove confirms by name and posts
brand-clear with no file.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Dropped into public/photos as home-hero.webp, which is all it takes: the build manifest picks up whatever is in that folder, and the hero already reads the home-hero slot. An upload in the panel still wins over it, so this is the default rather than a fixture. Converted from PNG to WebP, 423 KB down to 54 KB. It is the largest thing above the fold and is fetched at high priority, so the size is the difference between the page painting once and painting twice. Kept at its own 383x801 rather than upscaled in the file: the browser stretches it either way, and the extra bytes would buy no detail. Worth knowing what that means on a desktop. The picture is a portrait and the hero band is wide, so at 1280px it is stretched 3.3x across and only a slice of it is in frame -- it reads as a dark texture at 25% opacity rather than as a photograph. On a phone it fits almost exactly and reads as intended. A wider original, or the same shot framed landscape, would carry the desktop view too. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Site images: a grid that lines up, a frame you can crop and zoom, and a hero photograph
**The right-hand end of the footer is now Find us**: the address, a map of it, and a directions link. The social icons move up under the business name, which is where a reader looks for both anyway and leaves the last column free for the map. **The map is built here from a place name, not from a URL typed into the panel.** An <iframe src> is the one field where a pasted address is genuinely dangerous -- whatever it points at renders inside our page. Encoding a search term into a URL this code constructs means the panel can only ever move the pin, never change what is embedded. Leave the box empty and there is no map rather than a frame pointing nowhere. It is lazy, and proven to be: nothing is fetched until the footer is scrolled to. A map at the bottom of the page should cost nothing to anyone who never reaches it. **Website content gains a Footer section** -- the line under the business name, the heading above the map, the address one line at a time, where the map should point, and the directions link text. The blurb was written into the component, so the one paragraph describing the business could only be changed by a deploy. Verified: four columns on a desktop at 84, 370, 656 and 942px and stacked on a phone, the map 252x160 and 348x160 with loading="lazy", the address three lines, one directions link, the social icons in the first column, and nothing overflowing 390px. With the network stood in for, the frame is not fetched before the footer is scrolled to and paints once it is. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
"The images I added are not on the site" has two causes that look identical from a browser and need opposite fixes. Either the panel has not stored the picture -- nothing uploaded, or site_images never created on this database -- or it has, and the copy of the website published on the server was built before that slot existed and has nowhere to put it. Nothing anywhere said which, so the only way through was to guess and check. Everything in this repository is on the far side of that question already: the panel returns every slot it holds, and the built site asks for every slot it is given. Verified by serving the published build against a stub panel -- all sixteen pictures on the home page, and every one on the seven other pages, come from the panel when the panel answers. So the remaining failure lives between the two, and it is a publishing step rather than a fault. The site check now names it: - public-content.php reports imagesReady separately from the picture list. An empty list means "nobody has uploaded anything" or "the table does not exist", and those send someone to two different places. - publish-site.mjs writes build.json -- when the build was published, the commit it came from, and the hashed bundle names. A server holding an older publish says so instead of looking current. - fleet-check.html asks both, loads each stored picture to see that the file is really there, and searches the site's own bundle for each slot name. A picture stored for a slot the running code has never heard of is exactly a stale deploy, and it is invisible from every other angle: the panel is right, the file is there, and the page keeps showing the drawing. Checked against six servers -- healthy, no table, nothing uploaded, a stale deploy, a missing file, and no build.json -- and each one reports its own cause. The Pictures section in the panel also stops describing itself as the logo and the mark. It has covered every picture on the site since the slots were added, and a heading that says otherwise is why someone would not think to look there. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Footer map, a Footer section in the panel, and a check that says why an uploaded picture is missing
Three things, all on the top of the home page. The uploaded background was invisible. Two separate dimmers sat on it: the image itself at 25% opacity, and a navy gradient over that reaching 95% in the middle. Multiplied, about one part in fifty of the picture survived -- so uploading a background looked exactly like the upload not saving, which is what it was reported as. The image is at full strength now and a single scrim carries the legibility, the same weighted one the page banners use. It needed its own variant rather than .page-scrim. Four of the hero's five blocks are a left-hand column, and then the three points at the bottom run the full width, so a scrim that clears on the right the way the banners do would leave the last of those three on bare photograph. This one clears less far and deepens towards the bottom where that row sits. Measured rather than judged, against a pure white photograph -- the worst case, and not a rare one, since a bright sky behind a car is the commonest thing an owner uploads. Every piece of text: 4.98:1 at the tightest, across 1536, 1280, 768 and 390. Below 1024 it goes to even coverage, because there every line runs the full width and there is nowhere the photograph can show without something being read over it. The band of empty navy between the menu and the first word is gone. The hero had even padding top and bottom, which is right for a section in the middle of a page and wrong for the one thing at the top that should be read without scrolling. Tight above, unchanged below. And the page is wider: 86rem rather than 72rem, with the side gutters growing at sm and lg instead of staying at 20px. On a 1536 screen the content ran 1152px with 192px of nothing on each side. Seventeen containers, all the same change, so the header, the sections and the footer still line up. No horizontal overflow at 320, 390, 768, 1024, 1280, 1536 or 1920 on six routes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Let the home page photograph show, and give the page its width back
…rives
The three highlights were full-width rows, each split into a description and
a bulleted panel. Three of them ran to about two and a half screens, and every
row left a column of empty navy beside whichever half had less to say -- which
is what the gap in the screenshot is. Side by side they are one screen, and
three offers next to each other can be compared, which is the only reason to
put three offers next to each other.
The inner panel goes with it: a bordered box inside a bordered box was what
made the emptiness look deliberate. The points sit under a gold hairline in
the same card now, and the button is pushed down with mt-auto so all three
line up however long the copy above them runs. They stagger in on scroll and
lift on hover -- .card-lift as the light cards already use, with a gold glow
instead of a drop shadow, because a dark shadow under a dark card is nothing.
--- and then the speed.
The page was served as an empty <div id="root"> with 365 KB of JavaScript
beside it. Nothing at all was on screen until that had been fetched, parsed
and run: three seconds on a mid-range phone on a rural connection, which is
most of this site's visitors, and everything a page is scored on happens
inside that window. Three changes, largest first.
The site is rendered to HTML at build time. vite build --ssr produces a server
bundle, and prerender-seo.mjs -- which already wrote each route's title and
description -- now also writes its markup. The script still loads and takes
over; it is what makes the fleet list and the enquiry form work. But the words
and the layout are in the file the server sends, so the page is readable while
it arrives. prerenderToNodeStream rather than renderToString, because the
routes suspend now and renderToString would bake the fallback into every page
but the home page.
main.tsx keeps createRoot rather than switching to hydrateRoot, deliberately.
The footer prints the current year and the date picker opens on today's month,
so a page built yesterday and read today would not match and React would tear
it down anyway, loudly. createRoot replaces the markup in the same frame it
paints, which costs one extra render and cannot be wrong.
Every route but the home page is now fetched when it is asked for. One bundle
held all fourteen, so someone landing on the home page downloaded the tariff
table, the blog, the places list and the enquiry form before the first word
appeared. Home stays eager: it is where most people arrive, and splitting it
would add a round trip before the thing they came for.
Poppins is served from this site instead of Google. A third-party stylesheet
is the slowest thing on a cold load -- the browser cannot paint until it has
it, and getting it costs a DNS lookup, a handshake and a round trip, then all
three again to fonts.gstatic.com for the files it names. The rupee sign was
its own tax: Google puts U+20B9 in the devanagari subset, so every page with a
price pulled 38 KB of Hindi per weight. Subset to that one glyph it is 570
bytes, and the four weights together are 2.3 KB instead of 155 KB.
The hero photograph is marked as the priority fetch, since it is what Google
measures this page's loading time by.
Measured on a throttled phone -- 390px, 1.6 Mbps, 150 ms latency, 4x CPU:
first paint largest paint
/ 3124 -> 1436 ms 3452 -> 2032 ms
/cars 2784 -> 1108 ms 3048 -> 1440 ms
/tariff 2760 -> 1112 ms 3024 -> 1448 ms
All thirteen routes checked after: no console errors, one header and one
footer each, no sideways scrolling. An upload in the panel still replaces the
committed photograph once the panel answers, and the committed one is what
goes into the HTML -- which is the right way round, since a crawler should not
wait on a request to /admin.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Three offers across, and a page that is readable before the script arrives
Every image on this site is stored above the document root and streamed back by a small PHP reader, so the address a visitor sees is one this code makes up. It made up photo.php?f=v12-9f3a1c4b7d2e5a60.jpg. On a business whose traffic comes from searching for "self drive car rental Nagercoil", the file name is one of the few things Google has to go on about a picture besides its alt text and the page around it -- and that one said nothing at all. They are named for their subject now: /admin/images/self-drive-car-rental-nagercoil-9f3a1c4b7d2e5a60.webp /admin/photos/maruti-swift-self-drive-car-rental-nagercoil-<hex>.jpg /admin/place-photos/vattakottai-fort-places-to-visit-kanyakumari-<hex>.jpg The words come from the site, never from the uploader. Each site-image slot has a phrase in SITE_IMAGE_KEYWORDS -- what is in the picture, not where it goes, because "Cars page banner" is of no use to anyone searching. A vehicle's photograph takes the car's own make and model; a place's takes the place's name. The random suffix stays: it is what makes a replaced picture a new address, which is what lets the reader cache for a year. A path rather than a query string, so the words are in the address and not only in a parameter. admin/.htaccess routes /admin/images/, /admin/photos/ and /admin/place-photos/ to the three readers, which still answer ?f= as well -- every address stored before today keeps working, and so does every file name, because the one pattern that replaces the three old ones matches v12-<hex>, p3-<hex> and bcars-hero-<hex> too. No migration, nothing to re-upload. The pattern is the security, not the rewrite. Lowercase words joined by single hyphens, sixteen hex, one extension: there is no way to write `..` or a separator in that. Checked both ways round -- ten kinds of hostile input reduced to harmless words (../../nitesha-config/config becomes nitesha-config-config), and twelve shapes that must never be answered, including name-<hex>.jpg.php and a traversal appended to a valid name. Two things found while testing it. admin/src/ had no deny rule, so its files were reachable over HTTP. Nothing leaks today -- every one of them only defines functions, so a request returns an empty 200 -- but that is a property of each file continuing to be written that way rather than a guarantee, and sql/ and tools/ were already denied. And the readers now answer a conditional request with 304 instead of the file. These addresses change whenever the picture does, so the year-long cache was already right; what was missing was the ETag that lets a returning visitor spend a few hundred bytes instead of a few hundred kilobytes. Alt text, which is the other half of this. The six service cards and the three steps passed none, so an uploaded photograph of a wedding car was announced to nobody and read by nothing. They pass the card's own title now. The drawings stay decorative, which is still correct -- SectionArt already makes that call. The site check tells the two new failure modes apart. If a picture will not load at its new address it tries the old one: if that works, the file is fine and the rewrite is not, and the page says so and names the remedy rather than reporting a missing picture. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Name every uploaded picture after what is in it
Two things: the site got a lot faster, and the block of structured data every
page carries now describes an actual business rather than a name and a phone
number.
--- speed
Nothing was compressed. That did not matter much while every page was an empty
<div id="root">; it matters a great deal now that each one carries its own
markup, because the home page went from 3 KB to 74 KB in the process. Gzipped
it is 13 KB. Brotli first where the host has it, gzip behind it, and images
and fonts deliberately left alone -- WebP, AVIF, JPEG and WOFF2 are already
compressed and running them through again spends CPU to make the file very
slightly larger.
Measured on a throttled phone -- 390px, 1.6 Mbps, 150 ms latency, 4x CPU:
first paint largest paint
/ 1484 -> 692 ms 2124 -> 1344 ms
/cars 1108 -> 652 ms 1356 -> 908 ms
/tariff 1108 -> 640 ms 1380 -> 916 ms
Against where this started two changes ago, the home page's first paint has
gone from 3124 ms to 692 ms.
One thing removed rather than added. The first version of this preloaded each
page's banner photograph by hand, and that was a second download of a file the
page already had: React emits a preload for an image marked fetchPriority=high
as it renders, with the exact URL the <img> uses, and mine used the absolute
form of the same address. Found by reading the built HTML rather than by
trusting it.
--- what the structured data says
It said the name, a phone number, a locality and a district. It now says:
the postal address, including the postcode, read out of the panel's own
footer fields rather than typed a second time for the crawler -- the two
disagreeing is exactly how a business loses Google's trust in both;
opening hours, from a new field in the panel, in the form schema.org wants,
shown in the footer as well when it is filled in;
every town served as its own City, not just "Kanyakumari district". Someone
searching for a car in Marthandam is searching for a town, and a district
named on its own does not say the town is covered. Twelve of them;
the social profiles, as sameAs, which is how a page claims an account rather
than merely linking to one. Only the ones that are filled in;
a stable @id, so thirteen pages describe one business rather than thirteen
businesses that share a name;
hasMap, pointing at the same place the footer's map does.
serviceArea is gone: areaServed now says the same thing town by town, and two
properties describing one thing with different precision is a pair to keep in
step for no gain.
Every inner page also carries a BreadcrumbList, so a result prints
"niteshacars.in > Wedding car rental" instead of a bare URL. The leaf is the
title cut at the business name or the place, because a breadcrumb is read at a
glance and the full title is not.
robots.txt stops sending crawlers into the panel -- a login page can only ever
rank for the business's own name and disappoint whoever clicks it -- while
still allowing the three image readers, because a photograph that cannot be
fetched cannot be in image search.
Checked: all thirteen routes render clean, the structured data parses on every
page, and exactly one preload for the banner rather than two.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Compress what is served, and tell Google what this business is
Somebody in Marthandam looking for a car does not search for "Kanyakumari
district" -- they search for Marthandam. The site named twelve towns in a row
of chips on the home page and competed properly for none of them, because a
page that mentions a town in passing has nothing to say about it.
Twelve town pages now, and a hub above them at /car-rental. Each one says what
that town is, what the roads around it are like, what people actually hire
there, and how far the drives worth making from it are.
The risk in doing this is the opposite failure, and it is worse than not doing
it: twelve pages that are one page with the name swapped. Google calls those
doorway pages and treats them as a reason to trust the whole site less. So
every page is written from its own entry in towns.json rather than generated
from a template with a placeholder in it -- Colachel is about a fishing
harbour and a battle in 1741, Marthandam about the rubber belt and the airport
run, Padmanabhapuram about a palace that shuts for a long lunch and catches
people out. Measured rather than assumed: the two most similar pages share 63%
of their vocabulary, and most of that is the header, the nav and the footer
that every page of the site carries.
Distances are approximate road distances and are in the data file, so they are
one edit each rather than a search through markup. They want checking against
what the owner actually drives -- that is the one part of this I cannot verify
from here, and it is flagged rather than left to be discovered.
Structure, because a set of pages nothing links to is a set of orphans:
the town names on the home page are links now rather than chips that do
nothing -- that section was the obvious place and it led nowhere;
every town page links to the other eleven, so a crawler that finds one finds
all of them;
"Where we deliver" is in the footer, which puts the hub on every page of the
site rather than only on the home page;
all thirteen are in the sitemap, derived from towns.json by the same module
the app reads, so adding a town cannot produce a page the sitemap does not
list.
Each town page also carries a Service node naming that one town, pointing at
the business's @id rather than repeating it, and a three-step breadcrumb --
Home > Where we deliver > Marthandam. The leaf was "Self Drive Car Rental" on
all twelve at first: the rule that shortens a title by cutting at " in " threw
away the only word that distinguished them.
An unknown slug renders the 404 page rather than an empty frame with a name in
it, which is what a route matching anything would otherwise turn every typo
into.
readablePhone moved out of the footer into lib/phone.ts. The town pages print
the number too, and the second copy of a formatter is the one that goes out of
step with the first.
Checked: all twenty-six routes render with one h1, one header and one footer,
no sideways scrolling at 390px, and no console errors.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
A page for each town we deliver to
Five boxes in the booking screens take a file: Payment Screenshot / Receipt,
Deposit Proof, Refund Proof, Pickup Photos, Return Photos. They have been
there since those screens were built and were wired to nothing at all. No code
read the input, there was no table to put a file in, and nothing anywhere
displayed one. Choosing a screenshot and pressing Save recorded the payment
and dropped the file on the floor, silently -- which from the outside is
indistinguishable from it having worked.
Three things were missing and all three are here now: a preview so you can see
what you picked, the upload itself, and somewhere on the booking for them to
appear afterwards.
booking_files, one row per file. kind says which box it came from, and
ref_id ties a payment screenshot to the payment it proves rather than to
the booking in general -- four payments would otherwise leave four
screenshots in a heap nobody can match up;
api/booking-files.php takes several files in one request, because the pickup
and return boxes accept several and five round trips over a phone
connection is how some of them go missing. A file refused for being the
wrong sort is reported without losing the others;
the detail screen shows each kind under the section it belongs to, click to
open full size, with a remove button on the corner of each.
The upload runs after the record it belongs to is saved, not with it. A
booking_files row points at a booking, and it also means a refused image never
costs someone the payment they just typed in: that is already saved, and the
file problem is reported on its own.
These are not like the other images on this site, and the reader is where that
shows. brand.php and photo.php are deliberately public -- a vehicle photograph
is an advertisement. A payment screenshot carries a customer's name, their
bank and an amount; a pickup photograph shows a registration plate. So
booking-file.php requires a signed-in user, sends Cache-Control: private,
no-store so nothing in the middle keeps a copy, and the names are random
rather than descriptive -- a file name should not read "advance-payment-10000".
A random name is not access control; the session is.
Verified rather than assumed. Signed out, the reader answers 302 and zero
bytes. Signed in, the image and no more. Signed in with a name that is not
ours -- a traversal, a .php, a .jpg.php, a short hex, an uppercase one -- 404
every time. In the panel: five inputs wired, two files chosen give two
previews with the right names, reopening the modal clears them, the detail
screen groups by kind and renders nothing for a kind with no files, and a
caption containing <img src=x onerror=...> comes out as text.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Keep the screenshots the panel has always asked for
Thirty-nine figures changed. They were written from recollection and flagged
as such; these are checked against the towns' own Wikipedia articles where
those state a distance, and against the road figures the travel sites agree on
where they do not.
Two were badly wrong rather than slightly off:
Kuzhithurai is 26 km from Nagercoil, not 35;
Boothapandi to Tirunelveli is about 70 km, not 60 -- Nagercoil to Tirunelveli
is 81 km by road, and Boothapandi is only 11 km up that road.
Azhagiapandipuram was out by nearly half: 15 km north of Nagercoil rather than
22. Colachel is 20 km, not 27. Marthandam is 26, not 30. Thuckalay 16, not 20.
Eraniel is further than stated rather than nearer -- 20 km, not 17.
The airport runs were both overstated, which on a page about hiring a car for
exactly that run is the number people would have checked: Marthandam to
Trivandrum airport is about 45 km rather than 55, and Kuzhithurai about 42
rather than 48.
Smaller corrections follow from the corrected bases: a trip between two towns
cannot be longer than going through Nagercoil when Nagercoil is on the way, and
several were. Marunthuvazh Malai is 11 km from Nagercoil and 5 from Suchindram,
both verified, and both were out.
One sentence had to go with them. Kuzhithurai's page opened "closer to
Thiruvananthapuram than to Nagercoil", which the corrected figures make false
-- 26 km against 42. A number that is wrong is checkable; a sentence that
sounds like local knowledge and is wrong is worse, because nobody checks it.
The point it was making, that the work up there points north, is made without
the claim.
Two notes were rewritten for the same reason: "the long way round the coast"
for a drive now routed through Nagercoil, and "under an hour and a half" for a
42 km run that takes about an hour.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Check every distance on the town pages, and fix the wrong ones
An audit of the three panel pages -- every tab walked, the DOM checked for the faults that make a control lie about what it does -- and the things it found. Twenty-two duplicate element ids on Website content. Each section posts its own form, so "heading" in one and "heading" in another are different fields with the same name, which is right for the form and wrong for the document. The consequence was not theoretical: <label for="f_f_heading_"> finds whichever element the browser reaches first, so clicking the label above one section's Heading box put the cursor in a different section's. Ids carry the section now. Checked by clicking every label that reads "Heading" and seeing where the focus lands: seven sections, seven correct fields. A booking whose total did not match the rental amount above it. The screen said Rental Amount 5,000 and Rental Amount Due 6,542 with nothing on it accounting for the 1,542, because extra KM is charged in the total and shown three sections further down. The Payments block adds up in the open now -- the agreed rental, the extra KM with the count, any other charges or discount, and then the amount due. The field at the top is "Agreed Rental", because two things called Rental Amount that differ is the confusion itself. A balance of -3,458. The minus is easy to miss on a phone and reads as money owed when it is money to give back. It says "₹3,458 to return" now, and "₹3,458 due" the other way, and "settled" at zero. install.php and session-check.php were reachable over HTTP. Neither is a way in -- the installer refuses to run once an account exists -- but both print server paths, the PHP build and the cookie settings, which is nobody's business. Denied in admin/.htaccess rather than deleted, because deleting them is not the fix it sounds like: this host publishes by cloning the repository into the web root, so a file removed by hand is back on the next deploy. The rule says which two lines to comment out to install on a new host. That correction is mine: I asked for them to be deleted from the server several times, and it would not have held. The panel's own thumbnails fall back to the older address. They go through /admin/images/<name> now, which needs the rewrite rule, and the panel is exactly where someone looks to see whether an upload worked -- a thumbnail that depends on a server setting is the worst place for one to be missing. If the new address fails, the image and the crop window both retry brand.php?f= once. The enquiry search box had a placeholder and no label. A placeholder vanishes the moment anyone types and a screen reader need not announce it. Also checked and left alone: every PHP file and every panel script parses, the money tests pass 35 of 35, and the two remaining "unlabelled" fields the audit flagged are wrapped in a label element, which the check does not see. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Clear what was actually wrong in the panel
Three things asked for, and one found on the way that was quietly wrong. --- somebody else's car The business does two things and the schema described one. Most bookings are our own car, where the whole rental is ours. Some are a car sourced from another owner for a customer, and there the rental is mostly passed on and what the business earns is a commission. Recording those the same way overstated revenue by the owner's share every time -- the kind of error that only shows up when the money does not match the bank. A car now says whose it is. Ours, or someone else's with the owner's name and phone, and a "Temporary" tick for one brought in for a hire or two so it can be found and retired rather than sitting in the fleet forever. Partner cars carry a badge on the card, because whose car it is decides what a booking on it earns and nobody should have to open a dialog to find out. A booking on such a car asks for the commission -- and only such a booking does. A box that does nothing on our own cars is a box somebody eventually types into, and a commission entered there would quietly shrink the revenue that booking reports. The detail screen then shows the three figures that matter: what the customer pays, what we keep, what is payable to the owner. The dashboard totals both. --- extra KM was counted twice Found while doing that. b.total already includes the extra-KM charge, and Net Revenue then added extraKmRevenue on top of it, so every booking that ran over its allowance was counted twice for the overage -- on the dashboard and in the Revenue Report. The comment above it claimed the opposite. The extra KM comes out of the rental figure now, so the two cards add up rather than overlap, and both sums work from what the business earns rather than from what the customer pays. --- deleting a cancelled booking Only cancelled ones, and only once no money is attached. A cancelled booking still holding a payment or a deposit is not finished -- it is a refund waiting to be made, and deleting it would take the record of the money owed with it. The endpoint refuses and says how much is holding it. Cancelling stays the ordinary ending. This is for the ones that should never have existed: a test, a duplicate, a booking taken against the wrong car. The charges, readings, attachments and the files on disk go with it; the audit entry stays, so what a booking said is gone but that it existed and who removed it is still readable. --- part payments These already worked -- Advance, Balance, Additional, Extra KM, as many as the booking needs -- so what was missing was not the recording but the expecting. A monthly hire paid half now and the rest in a few days had no date on it, and the second half was remembered by whoever took the booking and nobody else. There is an optional "Balance due by" on a booking now, shown beside the balance while anything is still owed. Checked: nineteen browser assertions over the owner fields, the commission box, the brokered booking's figures, when Delete is offered, and the revenue sums; fifteen more over the split itself, including a database that has not run the migration yet, where every car reads as ours and every total is what it was. The money tests still pass 35 of 35. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Commission on other owners' cars, deletable cancellations, part payments
Most of the nine sections asked for were already here -- payment history with proofs, deposit collection and refund with deductions, pickup and return readings, extra-KM maths, a seven-stage lifecycle. What was missing was the evidence: the licence that was checked, the state the car left in, what came back damaged, and what was charged for it. Those are what a dispute turns on, and they were on somebody's phone or nowhere. --- the customer A WhatsApp number, because the one that answers a call is often not the one that answers a message and WhatsApp is how this business talks. A licence expiry, because a licence that runs out mid-hire is the owner's liability and worth knowing before the keys move. An ID number. A customer type -- new, returning, corporate. An estimated KM, which is useful before a hire for quoting and after it for knowing whose estimates to trust. --- documents Driving licence, Aadhaar or ID, passport, and anything else. PDF as well as photographs, because a licence usually arrives as a PDF -- and a PDF is recognised by its first five bytes rather than by what the file was called, for the same reason the images are. Against the customer rather than the booking. A returning customer's licence is the same licence, and asking for it again every hire is how a file ends up attached to three bookings and missing from the fourth. Uploading a licence replaces the one on file rather than stacking a fourth copy nobody can pick between; "other" is the exception, being where everything else goes. An expiry that has passed is marked, in red, on the booking. These are the most private things the system holds -- a licence carries a name, an address, a date of birth and a photograph -- so the reader requires a signed-in user, sends no-store, and the file names carry nothing. Signed out it answers 302 and zero bytes; a traversal, a .php and a .pdf.php all answer 404. --- the handover check Thirteen items on both legs: exterior, interior, tyres, spare, tools, papers in the car, AC, lights, indicators, horn, mirrors, wipers, toolkit. Stored as JSON on the reading rather than as thirteen columns, because the list changes -- a business that starts hiring bikes wants a different one -- and a schema change per item is how a checklist stops being maintained. Unticked is stored as false rather than left out. "Not checked" and "checked and wrong" are different things and only the first is what a blank should mean. --- charges at return Cleaning, fuel, late return, other -- each its own line with its own amount and note, and each feeding the booking's total. Itemised rather than folded into other_charges, because "₹1,150 of other charges" is not something a customer accepts and not something anyone can explain a month later. An "other" charge with no note is refused, and refusing it does not take the fuel charge down with it. Raised after the reading is saved, not with it, so a rejected charge never costs someone the reading they just took. --- damage What was damaged, what it is thought to cost, which leg it was noticed at, notes and photographs. Deliberately not money in itself: an estimate is not a charge. Some damage is billed, some is held back from the deposit, some is absorbed -- and a record that assumed the first would make the other two wrong. A tick on the return form raises the charge; leaving it unticked records the damage and bills nothing. --- the lifecycle Not changed, and this is the one place I did not do as asked. The statuses already run Enquiry, Confirmed, Ready, Active, Returned, Completed, Cancelled. The extra ones wanted -- Advance Paid, Payment Pending, Deposit Pending -- are not stages a booking is moved through: they are facts about the money, true or false on their own, and the panel already derives and shows them. Making them statuses would mean somebody remembering to move a booking to "Payment Pending", and a status nobody remembers to set is worse than none. Checked: twenty-nine browser assertions over the new fields, the checklist round trip, the charges going up line by line, the refusal that spares the rest, and what the detail screen shows; the document validator against a real multipart upload, where a PHP script renamed .pdf is refused and a genuine PDF is not; the reader's session guard; and the earlier suites still passing -- 35 money tests, 15 commission-split tests, 19 feature tests. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
The records a rental business is asked for when something goes wrong
A tick says "done". None of these lists are things that have been done: they are what is included in a hire, what the monthly rate depends on, what to bring, what we arrange for a wedding. A dot says "item", which is what they are. Thirty-one of them across the site. The three offer cards had a gold disc with a check mark inside it, which was also the heaviest thing in a card whose job is to be read; the hero, the monthly page, the NRI page and the service pages had a bare gold check mark. All of them are the same bullet now, and it is the same bullet the areas list and the town pages were already using -- there were two bullets on this site and now there is one. Sized and positioned in CSS rather than with a margin at each call site, because the one thing a bullet has to get right is sitting on the optical centre of the first line, and that depends on the line height beside it. The hero's list is 14px text and the service pages' is 16px, and a fixed margin cannot be right for both. lh gets it exactly, with an em fallback that is close enough everywhere it is used. Measured: the dot's centre lands at 10px against a first line whose centre is 10px. The one check mark left is the "your enquiry was sent" mark on the contact form. That one genuinely does mean done. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Every form in the panel closed on a click anywhere outside its card. On the booking form that is twenty fields and several minutes of typing, and it went with nothing asked and nothing kept. The owner reported doing it two or three times in a row on one booking. Three separate ways in, all closed: A press has to start AND end on the grey. Selecting text in a field and releasing outside the card was being counted as a click on the backdrop, which is the version of this that happens while you are still writing. A form with something typed in it asks first -- "Leave this form? What you have filled in is lost", Keep writing against Discard it, with Keep writing holding the focus so the reflex press keeps the work. The form is still behind the question, so answering it leaves every field exactly as it was. Escape did the same thing as the click and now goes through the same question. A form nobody has touched still closes straight away. Opening the wrong one and clicking away is not worth a question, and asking there would teach people to dismiss the question without reading it -- which is what makes it useless on the day it matters. Whether anything was typed is the form's values against the ones it opened with, snapshotted when the modal is shown rather than at the nine places that show one -- the tenth is the one somebody adds later without knowing this exists. A form that pre-fills itself from an enquiry is therefore not "edited" before it has been touched. Nine modals: the booking, vehicle, payment, deposit, refund, service, pickup and return forms, and the reminder form in the panel shell. The expense form never had the click-away behaviour and is unchanged. CHECKED IN A REAL PANEL untouched form closes on a click away yes, nothing asked typed-in form asks instead of closing yes "Keep writing" leaves the form open name and phone both intact Escape asks the same way yes a drag out of a field does not close it yes, nothing asked "Discard it" closes it yes reopening starts clean yes the reminder form behaves the same yes, text kept And the six UI suites still pass in full: 143 assertions across bookings, enquiries, finance, brand, navigation and the main suite -- all of which open, fill and cancel these same modals. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
A click an inch wide of the form no longer empties it
EDIT, EVERYWHERE, WITHOUT OVERWRITING ANYTHING
A payment entered as 8,000 when 3,000 was taken could be voided and
typed in again, and that was the whole of it: no other section of a
booking could be changed at all. Every section has its own Edit now --
payments, the charges added later, the deposit, each refund, the pickup
and return records, damage, the customer's documents, and the three
cards on Overview.
None of them overwrites a figure. What each one does depends on what
the row is:
a payment or an expense gets an adjustment row beside it, which is
what those two endpoints have always done
a deposit, a refund, a charge or an odometer reading is superseded --
the wrong row is marked with the reason it was wrong and the right
one cites it, the way a corrected reading already worked
a damage estimate and a document's details are changed in place.
Neither is money and no total is summed from either, so a second row
describing the same dent would be a worse record than one line that
says what the dent is. The before and after are in the audit log.
They all ask the same way, in a dialog form that is new here: what it
is now, what it should be, and why. Asking an amount, a date, a method
and a reason as four questions one after another is four chances to
lose your place, and it hides the old figure at the moment you are
typing the new one.
A correction is not an addition, so these stay available on a booking
that has been completed -- a figure found to be wrong a week later has
to be fixable. Adding something new to a closed booking is still shut,
as it was.
A CORRECTION COULD ONLY EVER REVISE A FIGURE UPWARDS
api/payments.php and api/expenses.php have both said since they were
written that an adjustment is "positive if too little was recorded,
negative if too much". Both validate the amount with the rule every
other amount in the panel uses, which refuses a negative outright. So
an amount entered too high could not be corrected at all -- in a year
of the panel being live, that path has never worked.
Validator::money() takes a $signed flag now, used by those two
endpoints and nothing else, and formatINR writes a negative as "- 500"
with the sign in front of the rupee mark rather than hidden beside it.
WEBSITE CONTENT AND PLACES WERE OPEN TO EVERYONE SIGNED IN
Both screens guarded nothing but being signed in, so an Auditor -- the
read-only role -- could rewrite the home page, and so could the counter
staff. Four permissions now cover them, with seeing and changing kept
apart: an Auditor reads the screens with every control disabled, and
Staff and Accounts do not see them at all. A Super Admin can still hand
either permission to one person without changing their role.
Three smaller permission faults, found in the same pass:
Staff was shown Finance and Reports and, on opening them, told it
could not load. The navigation is built from what the account may
actually do now, and the dashboard asks only for the figures it may
show -- it was firing two refused requests on every load, each one
writing a permission_denied row into the audit log.
"Set a reminder" is enforced by the server and was missing from the
Users screen, so it could not be withheld from anybody. The bell
offered "+ Add" to an Auditor, who cannot use it.
The top bar read "Admin" whatever you were signed in as, three inches
from a sidebar badge saying Staff. It says the role now, which on a
phone is the only place it appears at all.
ONE THING IN THE TOOLING
tools/clear-enquiry-throttle.php printed the panel's "not configured"
HTML page and exited 0 when it could not find a database -- so the two
enquiry suites called it, saw nothing wrong, and then failed three
assertions later because the rate limit was never cleared. A command
line gets a one-line error and a non-zero exit, and both suites now say
so instead of swallowing it.
The UI suites also leaned on fixed pauses and fixed test data: a phone
number reused every run so the list went on showing the first run's
name, a vehicle booked on the same twelve dates until eleven of twelve
came back 409, and sleeps that were long enough only on a quiet
machine. They wait for the thing they are waiting for now.
CHECKED IN A REAL PANEL
every Edit driven in a browser payment, deposit, refund,
charge, damage, document,
pickup, return, the cards
a correction of -500 on a payment received 4,500 -> 4,000
a charge corrected 500 -> 900 total 5,300 -> 5,900
one row left, not two yes, the old one marked
fuel level and condition corrected 1/2 -> 1/4, note kept
the same figure is refused "that is the figure
already recorded"
Auditor on Website content 245 of 253 controls
disabled, a forced save 403
Auditor on Places no form, no row actions
Staff navigation no Finance, no Reports,
no website, no 403s at all
the dialog on a phone fits 390px, one column
test-money.php 46 test-api.sh 20
test-auth.php 60 test-bookings.sh 31
test-api-failure.php 7 test-ledger.sh 80
test-ui.js 11 test-expenses.sh 40
test-nav-ui.js 17 test-enquiries.sh 36
test-brand-ui.js 13 test-finance-ui.js 30
test-records-ui.js 46 test-booking-ui.js 54
test-enquiry-ui.js 27
All green, and run in sequence against one database rather than one
at a time, which is how the interference above was found.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Found while checking the whole site: the footer's "Tariff" and "Enquire",
the "More about the Swift" under every fleet card, the car names in the
rate table, and "All articles" at the foot of a blog post were all 20 to
23 pixels tall. WCAG 2.2 puts the floor at 24, and a thumb is nearer 45.
All four are standalone links rather than words in a sentence, so they
take the tap-target and tap-list helpers the footer's page list already
uses -- which only grow on a touch screen, and leave a link inside a
paragraph exactly the size of its words.
CHECKED, ALL 43 ROUTES
every standalone tap target is at least 24px
every page fits 390px with no sideways scroll
every page has its heading and full text with JavaScript off
unique title and description on each, canonical and og:title on each
one h1 each, no heading jumps, no duplicate ids
every image has alt text and dimensions
no console errors, no failed requests, no dead internal links
structured data parses on all 43
Phone (4x CPU, 150ms RTT, 1.6 Mbps), median of three. Measured on this
build before the four links grew; a min-height on four links cannot move
a paint or a long task, and the crawl and the tap check above were both
re-run afterwards:
FCP LCP CLS TBT KB
home 916 1684 0.000 295 296
cars 736 988 0.000 149 193
tariff 768 1004 0.000 126 251
contact 776 980 0.000 154 199
a town page 748 988 0.000 118 186
a car page 724 972 0.000 154 193
a blog post 744 976 0.000 119 186
Every LCP under 1.7s and every CLS zero. The home page's 295ms of
blocking time is one 260ms task at 584ms -- the bundle compiling and
React hydrating -- and it does not reach the visitor: the worst
interaction measured on the same throttled phone was 88ms to open the
menu, 80ms to open an FAQ, 24ms to focus the enquiry form, all well
inside the 200ms that counts as good.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Every section of a booking can be put right, and no door the panel will not open
The enquiry form answered a refusal with "Please correct the highlighted fields." and highlighted nothing. The reason appeared in small red type under the box, the box itself looked exactly as it had before, and on a phone the whole thing was usually below the fold -- so the message named a cue that was not there and pointed at a field the customer could not see. A refused box now carries a red border and a faint tint, aria-invalid so it is announced and not only seen, and aria-describedby tying it to its reason. The three go on together through one helper, so a field cannot end up with the colour and not the announcement. The cursor moves to the first box at fault and the page scrolls it to the middle, which puts the reason on screen and has a screen reader read it out. Pickup location and the message box could be refused for length and said so nowhere at all; they say so now. The server was answering one fault at a time: a wrong phone and a wrong email meant being told about the phone, fixing it, sending again, and only then hearing about the email. The checks it makes after the Validator are gathered now and go back together, so "fields" is plural because more than one can be marked. The date picker already drew a taken day in red. A date the server refuses is the same thing to the person reading the form, so it is drawn the same way rather than given a second look of its own. Verified against a local copy of the document root served the way Hostinger serves it -- site at /, panel at /admin: 27 checks on the highlight itself, and the 34-check form run and 36-check endpoint suite still green. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
is_temporary has marked a car brought in from another owner since the commission work. It showed as Temporary in the fleet so it could be found and retired afterwards, and that is all it did -- the website never knew about it. So the fleet page, the tariff table and the car list on both enquiry forms were offering somebody else's car as part of what we hire out: a photograph and a price that are not ours, for a car that is gone next week, winning enquiries nobody can answer by the time they are answered. api/public-vehicles.php now excludes it, and api/public-availability .php with it. A car the site does not list cannot be the one a visitor picked, so its diary answers a question nobody asked -- and worse, it was counted in the fleet, so a weekend with every advertised car out still looked open because a car nobody can see was free on it. Nothing changes inside the panel. It is still in the fleet list, it still takes a booking, payments, handover and return like any other car. It is the shop window it stays out of, not the business. Both filters are guarded on the column existing, the way the rest of public-vehicles.php is: this endpoint serves the live website before anyone opens the panel to run a migration, and naming a column that is not there yet empties the fleet rather than showing one car too many. The panel now says so where the decision is made -- the tickbox explains it keeps the car off the website, the fleet card carries a quiet "Not on the website" beside the Temporary badge, and the line under Vehicle Management no longer claims every available car shows on the site. tools/test-public.sh is new: the public endpoints had no coverage at all, and they are the only door the fleet goes out through. 21 checks on what they show and what they hold back -- the registration, the odometer, the owner's name. Both halves of this fix were confirmed by turning each off and watching the suite fail. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
A refused field now looks refused, and a borrowed car stays off the website
Body prose was set at five different line-heights depending on which section it sat in. Most paragraphs carry leading-relaxed at 1.625; the ones written without it fell back to the browser's own 1.5; and four sections had been given hand-picked values besides -- 1.55 on the benefit text, 1.6 on the route sentences, 1.65 on the delivery intro, 1.5 on the place-card blurbs. Nobody decided that. It is what happens when each section is written on its own day, and the result is that two paragraphs of the same size, a screen apart, set differently. So the leading is set once, on p and dd in @layer base, from --leading-relaxed -- the value fifty-three places already used, so the exceptions join the rule rather than the other way round. The four sections stop naming their own number. Three small notes on /monthly and /nri that took text-sm's tighter 1.43 are given the house value: they close a section, they are not captions. The base layer is what makes it hold. leading-snug and leading-tight still win where a line is deliberately tighter -- headings, and the eleven-pixel note under the hero form -- and the next paragraph anyone writes starts level without being told to, instead of landing at 1.5 again. Left alone: the date stamp on a blog post. It is one line that never wraps, so its leading is not visible, and it is metadata rather than a section's prose. Measured in the browser rather than read off the stylesheet. Before: six values across the pages. After: one, on 43 routes at both desktop and phone width, with nothing scrolling sideways and no page erroring. The four sections grew by 4 to 7 pixels, and the three-line clamp on the place cards still clamps. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Five leadings for the same job
Every other "read more" link on the site grew for touch a week ago -- the fleet's "More about the Swift", the footer's "Get directions", "Why hire from us". The blog index was missed, so its five "Read it" links stayed 20px tall on a phone where the rest are 44. One class, the same one the others use. 69x20 to 69x44. Found by crawling all 43 routes at phone and desktop width. The three links reported at desktop width are the deliberate scope of .tap-target, which only grows under (hover: none) or 700px -- a decision to treat this as a touch concern, not an oversight, and one to revisit on its own rather than inside this. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
On a phone the blog index asked for a twenty-pixel thumb
.tap-target and .tap-list only grew under (hover: none) or 700px, so
with a mouse the footer's "Get directions", the fleet's "More about
the Swift" and the blog's "Read it" stayed 20px tall -- a line of 14px
text with nothing around it. WCAG 2.5.8 asks 24px of a target under
any pointer, and any pointer includes a trackpad, a tremor, a laptop
on a train.
So 24px is the floor at every width now, and the touch sizes stay
where they were. That is the whole change: the three links go to 24
and the footer bar's to 28, which is four pixels and one pixel. The
alternative -- giving a mouse the 44px a thumb wants -- would have put
two centimetres of nothing under every footer link on every page to
satisfy a rule that asks for half of it.
Checked at both widths rather than reasoned about: 43 routes, desktop
and phone, no target under its floor, and the phone sizes unchanged at
40 and 44.
Three times the crawl was wrong before it was right, which is worth
recording because each one would have become a bad change:
- It skipped every display:inline element as "in a sentence". A
standalone nav link is inline too. Tightened to ask whether there
is other prose around it.
- Tightened, it then flagged the fleet sentence on the twelve town
pages -- "Maruti Suzuki Swift, Toyota Glanza Automatic, ... and
Toyota Innova" -- because each car name is wrapped in its own
span, and a span holding one link has no other text in it. It now
measures against the nearest block.
- It still flagged "or call +91 63749 42976", on a threshold of
eight characters of surrounding text. "or call" is seven. The
threshold is gone: any prose in the block means the link is in a
sentence, which is what 2.5.8 says and what padding that link
would spoil.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
A mouse is a pointer too
useSeo() keeps the tab title, description and canonical in step after a
client-side navigation, because navigation never fetches HTML again. To
know what each route is called it imported towns.json, models.json,
service-areas.json and posts.json and composed the list in the browser.
All four landed in the entry chunk. So every page of the site carried
every blog post's body, every car's luggage capacity and the prose for
twelve town pages -- about seventy kilobytes of object literal, run at
module top level before anything on screen could answer a tap -- to end
up with a path, a title and a sentence.
The list is composed at build time now, by scripts/seo-routes.mjs, from
exactly the same four files and the same four helpers. prerender-seo.mjs
reads what it writes instead of composing its own copy, so the titles
written into each page's HTML and the titles useSeo puts in the tab are
one list and cannot drift -- which is what the old comment in useSeo was
reaching for when it said a second list is a second list to forget.
entry chunk 374,166 -> 314,667 bytes (-59.5 KB, -16%)
Every page pays that, not just the home page.
What it did for blocking time, phone at CPU 4x and 1.6 Mbps, median of
eleven runs per arm, two builds measured the same way:
TBT longest task long tasks
before 278ms 152ms 6
after 256ms 144ms 5
Honest about the size of that: the runs scatter 208-388 before and
153-351 after, so a 22ms median is a real direction and not a precise
figure. The 59.5 KB is the certain part.
Verified the output did not move: across all 44 prerendered pages the
only lines that changed are the hashed asset filenames. Not one title or
description differs. A browser check confirms the other half -- clicking
Tariff, Blog and Contact retitles the tab and moves the canonical, which
is the whole job of the file this touched.
Two wrong turns on the way, both measured and both reverted rather than
shipped. content-visibility on the offscreen sections, which cut the
longest layout pass from 73ms to 41ms and did nothing for blocking time,
because the cost was never layout. And the scene drawings: stripping all
176 SVG nodes changed TBT by less than the noise. The page is not slow
because of what it draws. It was slow because of what it carried.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Every page shipped the blog to put a title in the tab
There was no document to send anybody. The panel takes PDFs in -- a licence arrives as one -- and Reports has a Print/PDF button for the owner's own figures, but nothing produced a page for the person who booked the car. A Confirmation button now sits beside Edit and Cancel on a booking. It prints one page: who it is for, which car and its registration, from when until when, what was agreed -- daily rate, KM included, the rate beyond that, deposit -- and what the money is: rental, other charges, discount, extra KM, extras, total, received, still to pay, deposit held. Then a line about bringing the licence, and who to call. Printed rather than generated as a file. The browser's own dialogue saves a PDF on every machine and phone the panel is used from, with no library to ship and nothing to install on Hostinger. It costs one extra step -- choose Save as PDF -- and it works the first time on a phone at a handover counter, which a server-side generator would not have done without a dependency and a deploy. The filename is what makes it sendable, and browsers take it from document.title, so the title becomes the booking number for the length of the dialogue and is put back after. A customer receives NSC-2026-0024.pdf and knows what it is before opening it. The letterhead reads what already exists rather than adding a third copy of the phone number: name, phone and email from the website's seo.json, the address from the panel's own Website content, which is where the owner already edits it. Anything missing is left off instead of guessed -- a confirmation with no email line is still a confirmation; one with last year's phone number is a customer who cannot reach anybody. One defect found by reading the output rather than the code. The discount printed as a plain 500, so the column did not reach the total: 7,500 + 600 + 500 + 2,600 + 1,200 is 12,400, and the total said 11,400. The total was right -- it already had the discount off -- so the line now prints it taken off, and a customer adding it up arrives where the document says. That is the kind of thing a handover argument is made of. Verified by rendering real A4 PDFs from the shipped functions and the shipped stylesheet, not from a copy written for the test: four cases -- typical, completed with extra KM and extras and a discount, one with no address or licence or registration at all, and one with names long enough to wrap -- each one page, no errors. A test reads the printed figures back and adds them to check they reach the printed total. The panel is hidden while the sheet prints and the page is put back whichever way the dialogue ends. No database in this container, so the live panel could not be driven; the money and crash-handling suites pass (53 checks). Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
A booking confirmation the customer can be given
The confirmation had nothing to point at if a customer disputes the fuel level or comes back a day late. It does now -- but the terms are written in the panel, not here. That is the whole design. These are the business's terms of hire, and a rental agreement is not something for this file to invent and print on a document a customer is handed. So Website content gains a "Booking terms (printed on confirmations)" section: a heading and a list, one line each. The note beside it says what is worth covering -- fuel, late return, what the deposit covers and when it comes back, who pays a fine, what to do after an accident -- and leaves the words to the person who has to stand behind them. Written none, and the confirmation simply has no terms section. That is the honest result of nobody having written any, and better than a document carrying rules the business never agreed to. It ships empty. They print numbered, in 9pt under the money, because "point four" is how somebody refers to one on the phone. Eight terms is what a real set runs to, and on the fullest booking -- extra KM, extras and a discount, so three more money lines -- that spilled onto a second page. A confirmation sent on WhatsApp wants to be one file and one page. Tightening the margins would have bought about seventy pixels of the eighty-one needed and left the page cramped. Instead "What was agreed" now sits beside "The money" rather than above it: four lines next to nine instead of four lines then nine. That recovered 154px, and it reads better -- the rate and the KM allowance are the terms the bill beside them is worked out from. Six cases rendered as real A4 PDFs, each one page: typical, completed with everything on it, bare with no optional field filled in, long names that wrap, and both of those again with eight terms. Three more check the empty state the panel ships in -- no terms written, the field present but empty, and rows containing only spaces -- none of which produces a terms section, a stray heading or an error. The money column still adds up to the printed total. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
The logo. The letterhead was the business name typed out, on a
document meant to look like it came from somewhere. It now carries
the logo already uploaded in Website content, beside the name, capped
by height so a wide mark and a square one both sit on the same line.
No logo uploaded, no image -- the name still says who sent it.
The colours. It was black on white, which was the right instinct for
ink and the wrong one for a document a customer recognises. Navy
headings now, and gold twice: under the title and above the total.
Twice reads as the brand; everywhere would read as a template. All of
it is text and borders, never a background, because a browser printing
a page leaves backgrounds out unless somebody ticks a box.
The status line. The top right read "Completed" on a page headed
Booking confirmation, which raises a question it does not answer. The
status is the panel's business. It is gone -- except for Cancelled,
which is the one state that has to be said out loud, because handing
somebody a confirmation for a booking that is off is the only way
this document could mislead. That prints as a red box beside the
number.
The deposit. It was a line under "What was agreed" and a line again
in the money column, which reads as being charged for it twice. It is
not part of the total, so it is no longer a row in the column that
adds up to one: it is a note under it, saying how much is held and
that it comes back. One place, and the arithmetic is unambiguous.
A fifth thing the testing turned up rather than the eye. The logo is
written into the page and print() is called in the same breath, so
whether the image had arrived was a race -- and it lost often enough
in the harness to show up as a missing logo in three of eight PDFs.
Browsers usually wait. "Usually" is a confirmation handed over with
no logo on it, which is the kind of fault nobody reports and everybody
notices. printConfirmation now waits for it, for at most a second and
a half, and prints without it rather than not at all.
Eight cases rendered as real A4 PDFs, each one page, each checked for
whether the logo actually decoded: typical, completed with everything,
bare, long names, terms on two of them, cancelled, and one with no
logo uploaded. The money column still adds to the printed total and
the three empty-terms states still produce no terms section.
The harness had two faults of its own, both fixed: it named two files
the same so one case overwrote the other, and it extracted functions
by matching "function name(" which stopped finding this one the
moment it became async -- silently, which is the worst way.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
The confirmation was a plain sheet of labelled rows. This is the layout from the two screenshots: a navy band carrying the logo, the business details and the hero photo, with the booking number set over a scrim on the photo; a gold-ruled title; the booking as four cards in a 2x2 grid, each headed by an icon on a gold disc; the terms below; and a closing line. The footer bar marked in the second screenshot is left out. Three things the design did not settle, decided here: The deposit is one figure in one place. It sits under "What was agreed", which is where the agreed amounts are, and the note beside the money only says why the total does not include it. Printing the number twice reads as being charged for it twice. The closing line follows the booking's state. "Please bring the original licence" contradicts the Cancelled badge on a cancelled booking, and asks a finished hire for a licence already brought, so cancelled and completed each get the line that is true of them. Images are awaited before the print dialogue opens, up to two seconds. Without it the logo raced the dialogue and was missing from three runs in eight. Printed with print-color-adjust: exact, or browsers drop every background and the band comes out white. The confirmation prints full-bleed through a named page, @page confirmation { margin: 0 }, which leaves the reports on their 14mm. Checked by rendering eight bookings to real A4 PDFs -- typical, completed with extra KM and a discount, bare, cancelled, no logo, long names that wrap, and both with and without terms. All eight come to one page with the logo decoded, the money column adds up to the total it prints, and an empty terms field prints no terms section at all. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
The closing strip sat wherever the text stopped, which on a booking with no terms was two thirds of the way up with a third of the page blank beneath it. The sheet is a flex column now and the strip takes margin-top:auto, so it lands under the last thing on a full page and at the bottom of the paper on a short one. The mark moves into the strip rather than floating loose above it. Anchored to the paper it had nowhere safe to sit once the strip claimed the bottom: high enough to clear the strip put it through the terms. In the strip it has a place of its own and cannot land on anything, and the booking with no logo uploaded just gets a strip without it. Also keeps the phone number whole. It has a space in it, and with the mark taking room at the right the line broke through the middle of it, leaving the customer half a number to dial. Re-rendered the same eight bookings: all one page, logo decoded, money column still adding up to the printed total. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Booking confirmation: terms, the supplied design, and a footer at the foot of the page
The district had twelve town pages and twenty towns worth having one. These are the eight with something to say that the existing pages do not: Thovalai on the Madurai road, Kulasekharam for the falls and the hill country, Karungal, Muttom for the lighthouse and the rock shore, Manavalakurichi, Mandaikadu for the temple and the March festival, Kaliyakkavilai at the Kerala border, and Thiruvithancode. Muttom was already named as a handover point in pickups.json and had no page of its own, which is the kind of gap that makes a site look thinner than the business is. Data only -- towns.json drives the route, the metadata, the sitemap entry and the prerendered HTML, so adding a town adds a page. 43 routes to 51. Each page is written for the place rather than templated from the last one: what the roads are like, when the market or the festival makes them impossible, where a small car is the easier thing. The build's own length guard caught two descriptions that would have been cut in a search result; both shortened. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Eight phrases from the owner, placed on the page each one belongs to rather than repeated across the site: self drive car rental in Nagercoil already the page title car rental in Nagercoil already, 14 pages self drive car rental in Kanyakumari already the page title car rental in Kanyakumari already self drive cars in Kanyakumari added, Kanyakumari page cheap car rental in Nagercoil added, tariff 7 seater car rental in Nagercoil added, tariff and Nagercoil automatic car rental in Nagercoil added, tariff and Nagercoil Not a keywords meta tag: Google stopped reading that in 2009, and the same ten phrases on every page is stuffing, which is a demotion risk rather than a ranking one. These are sentences a customer would want to read whether or not a search engine existed -- which car is actually cheapest and why fuel matters more than the rate, that there are two seven-seaters and one automatic so a day's notice is the difference between having one and not. The two new tariff questions land in the page's FAQPage structured data, so they are eligible to show as rich results rather than only as body text. Verified by reading the built HTML back: every phrase resolves to the page it should. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Eight more towns, and the owner's search phrases placed where they are true
1. A phone number now has a country code in front of it, picked from a list of thirteen -- India, the Gulf, and the English-speaking countries the NRI customers come home from -- and the digits are counted against that country. The field took free text before, so "9876543210", "98765 43210" and "976543210" all went in and only two of them could be dialled. A number that cannot be dialled is worse than a missing one: nobody finds out until the customer is late and the car is gone. Stored as "+91 9876543210", which fits the VARCHAR(20) the columns already are. Numbers saved before this read back as +91, because that is who was being booked. Checked in the browser so the person typing finds out while the field is under the cursor, and again on the server, which is what actually keeps a bad number out of the column. Applies to the booking phone, the WhatsApp number, the partner owner's phone and the new enquiry correction form. 2. The address is a textarea rather than a single-line input. An address has lines in it; the old field made people run the whole thing together with commas, and that is what printed on the confirmation. 3. An enquiry's details can be corrected. The website form is filled in on a phone in a hurry, and a digit goes missing or the pickup is written as something only the sender would recognise. The panel could read that and not fix it, so the correction went into a note while the wrong number stayed in the field the panel dials from. The endpoint writes the before and after into the audit log, which is the difference between correcting a record and quietly rewriting it. 4. The rental now follows the dates. Changing the return date already moved the days preview and left the money alone, so a customer who brought the car back three days early kept the full hire on the booking until somebody retyped the figure from memory. The rate card's band for the days now on the form is offered under the amount, with a button. Offered, not applied: a rate agreed off the card is a real thing and overwriting it silently would be worse than never suggesting anything. Tested what can be tested without a database. The PHP splitter and joiner against nine stored formats including the pre-code ones, a round trip, and the digit counts. The browser half in headless Chromium: loading a stored number into two fields, nine length cases across four countries, and that switching India to Singapore makes a correct number wrong on the spot. The rate bands against nine hire lengths including the early-return case and a car with no weekly rate. The rendered controls for layout at desktop and phone width. Not tested: anything that needs the database -- this container has no MariaDB. The save paths are syntax-checked only. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
Four corrections to the panel's entry forms
The overview ran to 1,436px against about 674px of usable height on the
owner's monitor -- twice the screen -- and the window scrolled, which
took the sidebar and the search box with it. Three changes, measured
rather than guessed.
The window no longer scrolls on a desktop. The shell is the height of
the screen and the page area inside it does the scrolling, so the
sidebar, the top bar and whatever is at the head of a panel stay put. A
phone keeps ordinary page scrolling: a fixed-height shell there fights
the address bar hiding and showing, and loses.
The overview is two columns. Counters on the left, because they are a
fixed block you scan; what needs doing on the right, because that is
the part that grows. Each column scrolls on its own, so the overview is
one screen whether there are two alerts or twenty. It was ten sections
stacked down a 1,180px strip with a third of the monitor empty beside
it; the width cap goes to 1,520px for the same reason the lists already
had it -- counters and lists are not a paragraph.
A density pass over the whole panel: padding, gaps, card and heading
sizes. Nothing changed colour or disappeared.
The lists fill the screen and scroll inside themselves. A fixed 72vh
left dead space under the table on a tall monitor and cut it off on a
short one; the panel, the view and the wrapper each hand their
remaining height down, so the table ends where the screen does and the
filters and pager stay on screen. Rows went from 63px to 42px.
Measured at 1536x730 and 1920x950 by rendering the real markup against
the real stylesheet:
overview 1,436px -> 674px, fits exactly, window does not scroll
bookings 4 rows and 60 rows: window still, pager on screen,
table scrolls inside, header sticky over the rows
a tall tab bottom still reachable at 1536x730, 1280x620, 390x780
Two of those checks failed first on the harness rather than the CSS --
forty cards in a grid fit in five rows and prove nothing about
scrolling -- and were rewritten before they passed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0128YzbhrfGdegUSc9RrARRf
The panel fits the screen, and the screen stops scrolling
This branch has not been deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.