Gmails refresh-token lå i klartekst i databasen. Det er ikke filen på serveren,
der er problemet, men kopierne af den: sikkerhedskopier kan hentes ned gennem
fladen, så hver eneste kopi, der er taget indtil nu, bar et levende token.
Derfor ligger nøglen uden for databasen — i miljøet, eller i en fil ved siden af
dataene med tilstand 0600. VAPID-parret er præcedens for en hemmelighed,
instansen selv laver og gemmer i basen, men det mønster dur ikke her: en nøgle i
samme fil som det, den låser, er ingen lås. Backup er VACUUM INTO af selve
databasen, så filen følger ikke med.
Kun de navngivne felter forsegles — refresh_token, access_token, password,
client_secret. Resten af klatten er det, man læser, når en postkasse driller:
hvilken mappe, hvilken etiket, hvilken adresse. At kryptere det med ville ikke
købe noget og ville koste muligheden for at kigge.
Ingen migration, og ingen der skal forbinde forfra: en værdi uden mærket læses
som sig selv, og næste skrivning lægger den forseglet på plads. Kolonnen
konverterer sig selv efterhånden som den bruges, i stedet for i én omgang, der
skal have ret om hver eneste række på én gang. Der er en prøve for netop det.
Fladen siger det på siden, hvor man henter kopien — for den dag nøglen betyder
noget, er den dag, man gendanner et andet sted og undrer sig over, at postkassen
er tavs. Runen er på v4; en ny variabel uden et versionsspring er usynlig.
Co-Authored-By: Claude Opus 5 <[email protected]>
Første del af IMAP-understøttelsen: selve klienten, som sit eget hjørne uden
tråde ud i noget bestående. Den logger ind over TLS, henter de flagede beskeder
nyere end et givet uid, og giver dem tilbage i samme form som Gmail-klienten —
afsender, emne, uddrag — så den, der senere kalder dem, kan få at vide hvor der
skal kigges og ikke hvordan.
Flagede og ikke ulæste: ulæst er en tilstand, ejeren stadig bruger, og en synk,
der brugte den op, ville tage noget fra ham. Det er samme begrundelse som stjerner
på Gmail-siden.
TLS er ikke valgfrit, og der er ingen kontakt til at slå det fra. Det her bærer et
kodeord, der åbner en hel postkasse, og en kontakt til at fjerne beskyttelsen er
en kontakt, der en dag bliver fundet allerede slået fra.
iCloud og enhver almindelig udbyder taler det her. Office 365 gør ikke længere —
Microsoft har lukket for kodeord på IMAP og vil have sit eget OAuth — så det er
ikke en vært, man kan nå herfra, og at lade som om ville kun fejle senere og
mindre tydeligt.
Co-Authored-By: Claude Opus 5 <[email protected]>
De to ord "Omdøb" og "Slet" lå oven på gruppens navn, hver gang musen var i
nærheden. De er væk fra sidebjælken. Omdøbning starter nu ved dobbeltklik på
rækken — eller F2, hvis man har den i fokus og ikke vil bruge musen. Sletning er
flyttet til gruppens egen side, hvor projekterne, den ikke tager med sig, står på
skærmen ovenover knappen.
Dobbeltklik på et navn, der også er et link, er ikke gratis. Første forsøg lod
det første klik navigere, og så åbnede siden mens omdøbningsformularen kom frem —
farveprikkerne flyttede sig ud under markøren, hvilket testen fangede som "element
is not stable". Navigationen venter derfor to tiendedele på et muligt andet klik.
Det er under grænsen for et dobbeltklik og over grænsen for at mærke det.
Co-Authored-By: Claude Opus 5 <[email protected]>
Serveren svarede 502 med en forklaring, og Cloudflare erstattede kroppen med sin
egen side. Bevist i produktion: handleren skrev 122 bytes JSON på 38 ms med
teksten "the panel refused: HTTP 401", og browseren fik en HTML-fejlside. Det
gjaldt også Gmail, som bruger samme hjælper — derfor har fladen hele tiden kunnet
sige "noget gik galt" og ikke andet, selv om serveren vidste præcis hvad der var
sket.
En proxy har lov til at gøre det ved en origin-5xx. Så det er ikke Cloudflare,
der skal laves om: 502 var den pæne kode, ikke den brugbare. Upstream-afvisninger
er nu 422, som går uændret igennem. Navnet StatusUpstreamRefused står ét sted med
grunden, så det ikke bliver rullet tilbage af en, der med rette synes 502 lyder
mere rigtigt.
Fladen læste altid fejlkoden i kroppen og ikke statuskoden, så der var intet at
rette dér.
Co-Authored-By: Claude Opus 5 <[email protected]>
To fejl i samme række. "Omdøb" og "Slet" lå i flowet og optog deres bredde også
usynlige, så ARBEJDE blev tegnet som "ARBE"; de ligger nu over rækkens højre
ende og dukker op ved hover. Og chevronen sad før navnet og skubbede en gruppe
toogtyve pixels længere ind end de projekter, den indeholder — så en gruppe så
ud til at ligge under dem. Den deler nu plads med gruppens prik: prikken hviler,
chevronen tager over ved hover, og bliver stående mens gruppen er foldet.
De sidste otte pixels sad i a-reglen, som giver hvert link i sidebjælken sin
egen venstrepolstring oven i rækkens.
Prøven, der holdt øje med rækken, målte at alt lå side om side — hvilket var
netop den geometri, der skulle laves om. Den måler nu det, der blev bedt om:
at navnet kan læses helt, og at gruppens prik står præcis hvor et projekts gør.
Co-Authored-By: Claude Opus 5 <[email protected]>
Beskeden efter "Fetch now" stod på dansk uanset sprog, og med et "(r)" i stedet
for rigtig bøjning. Den bruger nu t() og plural() som resten.
Vagten burde have fanget den og gjorde det ikke: den krævede et anførselstegn
lige efter app.toast(, og denne stod som en skabelonstreng inde i en ternær hen
over to linjer. Den læser nu hele argumentlisten frem til den matchende
slutparentes, så det er formen på kaldet og ikke tilfældig placering, der afgør,
om en streng bliver set. Sat på prøve ved at lægge den gamle tekst tilbage:
begge strenge blev fanget.
Co-Authored-By: Claude Opus 5 <[email protected]>
Et 502 i produktion så ud som om kaldet til panelet hang. Det gør det ikke: et
panel, der aldrig svarer, giver 202 efter ti sekunder, og et panel, der afviser,
giver 502 med sin egen statuskode i teksten. Begge dele står nu som prøver, så
næste gang mistanken falder på handleren, kan den frikendes uden at gætte.
Hvad der så gav Cloudflares 502 i produktion, er stadig uafklaret.
Co-Authored-By: Claude Opus 5 <[email protected]>
En postkasse, der ikke svarer, holdt hele forespørgslen åben. Efter hundrede
sekunder gav Cloudflare op og sendte sin egen 502-side tilbage — HTML uden
fejlkode, så fladen kunne kun sige "Noget gik galt", og fejlloggen var tom,
fordi handleren aldrig nåede at vende tilbage.
Synk får nu sit eget budget på femogtyve sekunder. Løkken tjekker fristen mellem
hver besked, og hver opgave gemmes efterhånden som den laves, så en overskredet
frist er en kortere kørsel og ikke en tabt. Baggrundsjobbet beholder den fulde
tid; det har ingen, der venter.
Endepunkterne kan nu pege et andet sted hen, så testen kan stille en postkasse
op, der aldrig svarer, og se hvor lang tid der går. Budgettet kan sættes, så
den prøve koster et sekund frem for femogtyve.
Et budget på nul betyder standarden — ikke ingen tid. Uden det ville en
opsætning, der ikke satte feltet, lade fristen udløbe før første kald og gøre
hver eneste synk til et stilfærdigt nul.
Co-Authored-By: Claude Opus 5 <[email protected]>
Reported from use: "Fetch now" said "something went wrong" and the error log was
empty. Both halves were the same bug wearing two hats.
`handleGmailSyncNow` answered 502 with the code `internal_error`, which the
interface maps to a generic sentence — so Google's own words, which are the whole
diagnosis, were replaced by one that says nothing. And `writeError` records
nothing; only `s.internal` does. So there was no message to read and nowhere to
look it up afterwards.
`s.upstream()` is the answer: 502 because this is somebody else's service saying
no rather than this server breaking, a code of its own so the interface can quote
what was said, and a row in the error log — which is the entire reason that log
exists. A Gmail sync failing quietly for a week is otherwise invisible: the
background job writes a line the next restart erases.
The same connection also left the mailbox unnamed. The address is learned at
connect time and that call can fail; a warning went to the log and an empty string
to the database, so the page said "an unknown account" and kept saying it. The only
way back was to disconnect and reconnect, which nobody would guess. The status
endpoint asks again when it finds none, so a wrong answer lasts until the next look
rather than until somebody reconnects.
Ten Danish strings were hiding in `confirm()` and `app.toast()` calls, including
every destructive confirmation in the app — the worst possible set to leave in one
language, since it is what somebody reads in the half-second before agreeing to
lose something. The i18n guard reads markup and attributes, and a string argument
lives in the script block. It reads those now, and found four my own grep missed.
Co-Authored-By: Claude Opus 5 <[email protected]>
The list this session started with is empty, and so is the one that grew during it
— eleven items reported by somebody using the app while it was being built, taken
in the order they arrived.
Two traps worth the next session's time, both found the hard way. A Rune's manifest
`version` has to be bumped when its variables change: the panel keys its catalogue
on that number, so new fields with an unchanged version are invisible, the upload
succeeds, and nothing reports it. And the restart button's `confirm()` cannot be
driven from automation — a browser modal blocks the renderer — which is the point,
but it means an agent testing it has to ask a person to press OK.
Three things are reported and not fixed: Gmail showing "an unknown account" because
the profile call fails into a log line that a restart erases, an unidentified
"something went wrong" on the integrations page, and three Danish strings in a
script block that the i18n guard cannot see because it only reads markup.
"Det næste, jeg selv ville tage" was three items that are now done. It says what is
actually next.
Co-Authored-By: Claude Opus 5 <[email protected]>
The three panel settings were added to the manifest in v0.12.0 and `version` was
left at 2. The panel keys its catalogue on that number, so a manifest with new
fields and an unchanged version is one it has no reason to re-read — the server's
settings page simply does not offer them, and nothing reports it.
Co-Authored-By: Claude Opus 5 <[email protected]>
A container cannot replace its own image from the inside, so "update" is really
"ask whoever owns the container to recreate it" — Yggdrasil pulls `:latest` on the
way, which is what makes a restart an update at all. This removes the second
browser tab, not the panel's authority: the panel still does the work and still
records who asked.
Three settings, needed together, and the page names whichever are missing — an
operator reading "not configured" should not have to go and find out which of
three. The token is never sent to a browser; the page is told only whether it is
set.
Answers 202 rather than waiting. The panel stops this container as part of
answering, so the reply may never arrive, and a request cut off mid-flight is the
successful case wearing a failure's clothes. The interface believes neither answer
and polls /healthz until it responds.
Administrators only, and sessions only. A leaked API token that could restart the
server is a denial of service in one request.
Co-Authored-By: Claude Opus 5 <[email protected]>
Two from use.
Section headings were small uppercase text over a hairline — a heading you have to
look for. A project with four sections is four groups, and the eye should be able
to count them without reading a word, so the heading gets a ground of its own and
the rows sit against the page. `--surface` rather than a tint of the accent: the
colour on that page belongs to the project, and a section is not a second thing to
colour. A step in lightness separates them in all five themes without anybody
measuring anything.
And the Inbox. The server names it at account creation, before anybody has seen a
settings page, so it is the one project name nobody chose — and an account made in
Danish went on saying "Indbakke" inside an English interface, which reads as a
string somebody forgot to translate. It was. `projectName()` translates it while it
still carries the name the server gave it; rename it and that is your name for it,
in every language, like any other project. The rename field keeps the raw name, or
editing a translated label would save the translation.
The screenshot set gains the list view, which is where sections read as bands
rather than as columns — and the shape most people actually work in.
Co-Authored-By: Claude Opus 5 <[email protected]>
A password and a TOTP code are both shared secrets: the server keeps something
that can be stolen from it, and the person keeps something that can be talked out
of them. A passkey is neither. The private half never leaves the device, and the
signature is bound to the origin that asked for it — so a convincing copy of the
sign-in page at another address gets a signature it cannot use.
Logging in asks for no email. The device knows which account its key belongs to,
which is not only convenience: it means the sign-in page cannot be used to find out
who has an account here. A key that verified its owner is both factors at once and
completes the login; one that only proved possession still meets the TOTP step,
exactly as a password does.
The relying party id is derived from VERDANDE_BASE_URL rather than configured
separately — a second setting that has to agree with the first is one that will
one day disagree, and the failure is a login nobody can complete. An IP address is
refused up front: no browser accepts one, and without the check the server issues
a challenge happily and the browser throws SecurityError long afterwards. The list
endpoint reports whether this deployment can offer them at all, so the interface
never shows a button that cannot work.
A challenge is stored server-side, answerable exactly once, and taken rather than
read — two requests racing the same one would both succeed, which is the shape of
a replay. A sign count that has gone backwards means the credential was cloned and
is refused rather than logged.
The Go tests cover who may reach these endpoints and that challenges are spent.
The ceremony itself needs an authenticator, so the smoke test drives Chromium's
virtual one over CDP — the only thing that can tell whether the base64url
conversions are right, since one wrong byte produces a signature that will not
verify and reads as a broken key rather than a broken encoder.
That test runs at localhost in its own context while the suite stays on the
loopback address, because WebAuthn refuses an IP and resolving localhost is slow
enough here to cost the suite minutes.
Co-Authored-By: Claude Opus 5 <[email protected]>
It was created alongside the three untranslated strings it catches and left
untracked — `git add -u` stages tracked files only. CI has been running without
it since.
Co-Authored-By: Claude Opus 5 <[email protected]>
Reported from use: "I just closed a task by accident, and there is no undo and no
list of closed tasks." Both halves were true.
Completing is one click on a small circle beside a row you were only reading, and
it takes the row off the screen — so the mistake and the evidence of it leave
together. There was no name left to search for and nowhere to look: the project
page can show its own closed tasks, but only if you already know which project it
was in.
The toast covers the mistake you notice at once. It names the task, so you can see
which one you hit, and stands for twelve seconds rather than five — five is enough
to read "could not save" and not enough to notice, decide and reach the mouse. A
toast with an action also gets two targets: the message dismisses and the action
acts, which were one button before and would otherwise put "yes, put it back" and
"go away" in the same place.
`/faerdige` covers the one you notice tomorrow. A place rather than a toggle in
every view, because somebody hunting for a task they closed does not know which
view they closed it in. Newest first — a record of what is finished is read
backwards from now, not by a due date the task no longer has — and grouped by the
day, because "I dag" and "I går" is how anybody would describe it.
A recurring task is offered no undo. Completing it moves it to its next date rather
than closing it, so there is nothing to reverse, and an undo that quietly rolled
the date back would be a different act from the one it claims to be.
Co-Authored-By: Claude Opus 5 <[email protected]>
The box at the top puts things in the project with no section, which is right for
capturing. But a project has sections because the work belongs somewhere, and
dragging every new task down into the right one is a second gesture for something
you already knew when you typed it.
Two ways in, because they answer different moments. A field at the foot of each
section and each board column — standing there is enough, no syntax to know. And
`/section` in quick add, for the box at the top and for Today and Upcoming, where
you are not standing in a section at all.
The field goes through quick add rather than createTask, so a line written there
still parses its dates, priorities and labels.
The sigil needed care: '#' and '@' never turn up in a Danish task line by accident
and '/' turns up constantly. It has to follow whitespace and be followed
immediately by the name, which "km/t", "3/4" and "20 kr / stk" all fail. The bare
form is greedy, because "Kunder/Ordrer" is a real section name and the sigil is
inside it.
A section id from the client is checked against the project the task actually
landed in. A `#project` in the same line can move the task elsewhere, and the id
would otherwise follow it into a project that section does not belong to.
Co-Authored-By: Claude Opus 5 <[email protected]>
"GarageRisteriet" rendered as "GarageRist / eriet" — broken mid-word. A name split
mid-syllable does not read as a wrap, it reads as a rendering fault.
Two causes. The row carried eight controls beside the heading — colour, show-done,
three view buttons, share, history, delete — and the heading was the only flexible
thing in it, so it was what gave way. And `overflow-wrap: anywhere` breaks at the
first opportunity rather than the last necessary one, so what gave way, gave way
inside a word.
Everything but the view switcher is behind one button now. What stays out is what
you use while working — which view you are in. What moves in is what you do *to* a
project rather than in it, and most of it once ever. The heading gets the row back,
and `break-word` only splits a word that could not fit on a line of its own.
The menu closes on Escape and on a click anywhere else, listened for in the capture
phase so a click whose own handler stops propagation still closes it. A menu you
can only leave by pressing its own button again is one you click twice to get out
of.
Three more untranslated strings turned up on the way: "Fjern {member.name}", "Luk
menu" and "En slettet konto". The guard missed them because none of its words
covered them — it knows fjern, opret, vælg, luk, tilføj, ryd, gem, slet and omdøb
now.
Co-Authored-By: Claude Opus 5 <[email protected]>
Spotted on the running instance: the version line read "Running v0.11.0. Der
bliver ikke set efter opdateringer." — half translated, on one line.
The sweep that was supposed to catch these grepped for Danish letters, and none of
the three survivors contain one: "Ingen nyere version fundet.", "Ikke mere i dag."
and a token hint split across two source lines so neither half looked like prose.
So there is a test for it now. It reads the templates and looks for the words that
give Danish away whatever letters they are spelled with — a word list rather than a
language detector, because this ships as one static binary and thirty common words
catch the sentences a person actually writes. It fails on the exact three that got
through.
Co-Authored-By: Claude Opus 5 <[email protected]>
Reported as "sections can only have one task", and that is exactly what it looked
like.
TaskList read the dragged task from its own `draggingId`, which is only set when
the drag *started* in that list. A task dragged in from the unsectioned rows above,
or from another section, arrived as null and the handler returned early — after it
had already called `stopPropagation`, so the section around it never got the drop
either.
So a section could only be dropped into while it was empty, because an empty one
has no row in the way to swallow the drop. Put one task in and there was no route
to a second. The id comes out of `dataTransfer` at drop time now, and `dragover`
decides with `carries()` rather than local state — which is the only thing a target
can read while a drag is in the air.
The stuck frame was the same cause: the section's own cleanup never ran, so the
highlight stayed lit after the drop. It is cleared on `dragend` now, which fires
once per drag on the source and bubbles to the window — the one place that is true
however a drag ends, including one abandoned with Escape, which produces no drop at
all and left the frame standing too.
The smoke test drops a second task onto the row already in the section, which is
the gesture that did nothing, and runs in WebKit as well.
Co-Authored-By: Claude Opus 5 <[email protected]>
The `webkit-forms` project was added to the Playwright config and CI was not told,
so it installed Chromium and the one WebKit test died on a missing executable —
26 passed, 1 failed, and the failure said nothing about the app.
It is one test on purpose. The suite stays Chromium, because running it in three
engines finds the same breakage three times; focus is the exception, and Safari
ignoring `autofocus` on anything inserted after load is what made "sections have
no function" true from inside Safari.
Co-Authored-By: Claude Opus 5 <[email protected]>
The note on 0009 and 0012 said the drop would cascade and take every row with it.
It would not: nothing references `activity` or `attachments`, so dropping either
takes nothing along. That claim was carried over from 0008, where seven tables do
reference `tasks` and it is exactly right.
The real reason is the check that comes with the marker. With foreign keys on, the
copy is validated a row at a time against a half-rebuilt schema, and the first bad
row aborts with no idea how many others are like it. With them off,
`PRAGMA foreign_key_check` runs once before the commit and names what is actually
wrong — which matters, because these run against a live database on somebody's
next restart and a failed migration is a server that does not start.
The upgrade test now carries activity rows, an attachment and a group across all
five migrations rather than stopping at 0008, and checks that the columns added by
ALTER arrived with their defaults rather than as NULL.
Co-Authored-By: Claude Opus 5 <[email protected]>
A task list is what is left to do, so closing something has to take it off the
plan. But "what did I get done" is a real question and the answer was nowhere:
once a task was closed it left the interface entirely.
They come back in one list at the bottom rather than back among the open ones in
their sections. A closed task has stopped being work and started being a record,
and putting the record in the middle of the plan makes the plan longer without
making it say more.
Off by default, and deliberately: a list that opens full of things already done is
a list you have to read past before you can use it.
In localStorage rather than on the account, like the sidebar's width and unlike a
folded group. It is a way of *looking*, not a statement about the work — somebody
reviewing their week wants it on for ten minutes, and a setting that followed them
to every device is one they would have to turn off twice.
`loadTasks` asks for closed tasks when the setting is on, rather than each caller
deciding. A view added later cannot then forget and quietly ignore it.
Co-Authored-By: Claude Opus 5 <[email protected]>
The account has had a `locale` since the beginning and it only chose which grammar
quick add read a line with — so picking English changed nothing anybody could see,
and the field said so honestly: "Sprog i hurtig tilføjelse". It says "Sprog" now,
because it means it.
Every component goes through `t()`. Danish is the source: a key missing from
English falls through to Danish rather than to the key, because a Danish sentence
in an English screen can be read and `settings.notifications.title` cannot.
`tag()` is the single place `Intl` is asked. Before, `'da-DK'` was written into a
dozen components, and that is the half of a translation that gets left behind — an
English interface still saying "24. aug." is exactly the shape of a job half done.
There is a smoke test that switches to English and refuses to see a Danish month
name in the calendar's heading.
Colour and theme names are keys now. `colors.js` keeps its ids untouched — they
are what the database stores, and `internal/store` reads them out of that file —
but `name` is `'color.tomato'`, resolved where it is drawn.
The Inbox keeps its name. It is a project row the server creates at signup, not a
string in the interface, and it can be renamed like any other project.
Four copies of "3 hours ago" became one, in `web/src/lib/when.js`. They had
already drifted: one said "aldrig logget ind", another "aldrig", and two stopped
at days where the others carried on to a date.
Routes stay Danish. `/indstillinger`, `/projekt`, `/uddelegeret` are bookmarks and
they are in every smoke test; translating them buys nothing and costs both.
Co-Authored-By: Claude Opus 5 <[email protected]>
The account has had a `locale` since the beginning, and it only chose which
grammar quick add parsed a line with — so setting it to English changed nothing
anybody could see. This is the other half of that field, started.
`t()` reads `i18n.locale`, which is `$state`, so every component that calls it
redraws when the language changes; that is what makes the setting take effect
without a reload, and why the module is `.svelte.js`. Danish is the source and the
fallback: a key missing from English falls through to Danish rather than to the
key, because a Danish sentence in an English screen can be read and
`settings.notifications.title` cannot. `tag()` is the single place `Intl` is
asked, so dates follow the language instead of being `da-DK` in a dozen files.
`internal/i18n` is a Go package that exists only to fail the build when the two
dictionaries drift — the same trick `colors_test.go` plays on the palette. It
checks the placeholders as well as the keys: a translation that spells `{n}` as
`{antal}` leaves `{antal}` on the screen, which is worse than an untranslated
sentence because it looks like a bug in the app rather than a gap in a dictionary.
Converted so far: the shell. Sidebar, layout, quick add, task row, command
palette — fifty-two keys, and the surfaces people look at constantly. Twenty
files remain, the settings pages being the heavy ones.
The language picker is deliberately not wired up yet. It still says "Sprog i
hurtig tilføjelse" and still only changes the parser. Wiring it before the sweep
is done would hand anybody who picks English a bilingual screen — the
half-finished-theme failure `app.css` already taught us. As it stands Danish is
unchanged and English is not offered, which is the honest state for something
half done.
Co-Authored-By: Claude Opus 5 <[email protected]>
It was a heading over some rows: a name, a colour, and whether it was folded.
Enough to tidy a list, and not enough to be what people mean by it — "Arbejde" is
a body of work with context of its own, and documents that belong to all of it
rather than to any one project inside it.
`/gruppe/{id}` carries three things a heading cannot. The projects in it, each
with what is still open — open tasks, not all of them, because "12" beside a
project somebody finished last year is a number that means nothing. A description,
written in place and saved on blur, the same shape a task's description has. And
the files that govern the whole group: the contract, the drawing, the agreement
everything else rests on.
The name in the sidebar is a link to it and the chevron beside it still folds.
Two intentions, two targets — the single one they used to share could only ever
serve the smaller.
Migration 0012 rebuilds `attachments`, because its CHECK spelled out exactly which
parent columns may be set and SQLite cannot alter a CHECK. The rule is still
"exactly one parent", counted rather than written out as pairwise comparisons:
with three columns that is three clauses, with four it is six, and the version
that is wrong looks exactly like the one that is right.
A group's file is reachable by its owner and nobody else. A group belongs to one
person and is never shared, so `reachAttachment` asks about ownership rather than
a project role — both old callers resolved straight to a task, so a group's
attachment answered 404 for the person who uploaded it.
Three tests also stopped racing their own navigation. Creating a project navigates
and `goto` is a promise nobody was holding, so a quick add fired straight after
could land in the Inbox instead — intermittent, and it read as the quick add
having failed.
Co-Authored-By: Claude Opus 5 <[email protected]>
The OpenAPI document gains the audit log, the backups, the Gmail client and the
sidebar sections, and loses the paragraph that said deleting an account takes
every task the person ever wrote.
The handover gains the day: why an account can be deleted without taking other
people's work, why the audit log pages with a keyset, why a week view exists at
all, why form fields are 16px on touch, and why `autofocus` is gone.
And the trap that cost the most, which was not in the code: the build hung on 0%
CPU because iCloud had evicted node_modules and a stalled download blocks a read
forever. The fix is a repo path outside a synced folder.
Co-Authored-By: Claude Opus 5 <[email protected]>
They were many rounds behind — no settings, no task detail, no themes, no groups,
no colours, no calendar, no admin pages. The reason is in how they were made:
`tools/shots` needs a running instance and a session cookie pasted in by hand, so
it ran about as often as somebody felt like doing that.
`node scripts/generate-screenshots.js` starts the real binary, creates its own
account, fills it with Danish demo data and photographs eleven surfaces into both
`site/screenshots` and `docs/screenshots`. Nothing to set up and nothing to
remember.
A script rather than a test, for the same reason `generate-icons.js` is one: CI
must never rewrite a tracked binary file.
`tools/shots` stays. It is still the way to photograph an instance that is
already running rather than a fresh one.
Co-Authored-By: Claude Opus 5 <[email protected]>
The field was declared on the request type, commented on it, and applied by the
update handler — and the create handler never read it. Creating a project into a
group answered 201 and put it nowhere. A field that is accepted and discarded is
worse than one that is refused: nothing reports it.
Found by looking at a screenshot. Both groups stood empty while the projects sat
loose above them.
Co-Authored-By: Claude Opus 5 <[email protected]>
The one click that opens Google's consent screen already existed. What it needs
behind it is a registered client, and no app can conjure one: Google issues no
Gmail access to an unregistered client, and `gmail.readonly` is a restricted
scope, so an id shipped inside a public image would be both extractable from it
and unusable without Google's security review of every deployment using it.
The registration is a one-off. Having to edit a Rune's manifest and recreate the
container to paste the result was not.
`VERDANDE_GMAIL_CLIENT_ID` and `_SECRET` still win when set. A value in the
environment is a deliberate deployment decision, applied again on every recreate,
and a form that could quietly override it would make the manifest a lie — so the
page says where the value came from, and PUT answers 409.
The redirect URI is spelled out on the page rather than left to be reconstructed.
It is derived from VERDANDE_BASE_URL, it has to match Google's copy exactly, and
the error Google returns when it does not names neither value.
Its own table rather than `user_settings`: an OAuth client is the instance's, not
one person's — and `user_settings` has `user_id` as its whole primary key, so a
second scope there would silently overwrite the AI provider's key.
Co-Authored-By: Claude Opus 5 <[email protected]>
A task with three sub-tasks looked exactly like one without, and so did a task
with a drawing on it. The whole reason to attach something is that somebody
should find it.
Counted in the same query as the tasks, as correlated subqueries: a list of fifty
would otherwise be a hundred more lookups, and the point is that you see this
without asking.
Said as a fraction rather than a total. "3" tells you something is there; "1/3"
tells you whether there is anything left to do, which is the question a list is
for. All-done goes accent-coloured, because "3/3" and "1/3" are the two answers
you actually look for and only one of them is good news.
Co-Authored-By: Claude Opus 5 <[email protected]>
Two reports, one sentence between them: "sections have no function, because you
cannot create more than one."
`autofocus` does not work in Safari for anything inserted after the page has
loaded, and every inline form here is inserted after the page has loaded — rename
a project, add a section, name a group. In Safari they all opened with no cursor
in them and what you typed went nowhere. That does not look like a bug. It looks
like the feature does not work. `use:focusOnMount` replaces all seventeen.
And the board had no way to add a section at all. The button lived only in the
list view, which is the one place nobody looks for it: a column on a board *is* a
section, so that is where you go to make one.
Chasing this turned up a third. The task drawer's three fields save on blur, and
closing the drawer *is* a blur — the input is being unmounted and the browser
fires blur on the way out. If the task left the store in the same moment, `task`
was already null and the handler threw `Cannot read properties of null`, stopping
the page. It only became reproducible once focus actually worked.
Fields are 16px on a touch device now, whatever the component asked for. Below
16px, iOS Safari zooms in when a field takes focus and does not zoom back out —
so the whole interface is left too large and has to be pinched back by hand, once
per field. The type scale tops out at 15px, so every input in the app did it.
The sidebar's resize handle sat at `right: -3px`. An absolutely positioned child
counts towards its container's scrollable overflow, and `overflow-y: auto`
resolves the other axis to `auto` too — so those three pixels gave the whole menu
a horizontal scrollbar on every desktop, always. The account name could do it as
well: `overflow: hidden` with no `min-width: 0` cannot shrink, so the row grew
instead of ellipsing.
There is a WebKit pass in the suite now, over one test. The rest stays Chromium
on purpose — a smoke test exists to catch a build that does not work, and three
engines find the same bug three times — but focus is the exception, and it took a
person using the app to find that out.
Co-Authored-By: Claude Opus 5 <[email protected]>
It has run since the beginning and nothing showed it. A backup nobody can see is
a backup nobody knows has been failing, and the failure is quiet: a volume gone
read-only, a disk that filled, a container that has not been up at midnight all
week. The first anybody hears of it is the day they need one.
`ListBackupRuns` was already in the store and exposed nowhere. Now there is a
list, a download, and a button that takes one now — the same work the nightly job
does without its "one a day" check, because waiting until tonight to learn whether
backups work at all is how somebody finds out on the day they need one, and it is
the only way to get a snapshot from *before* something you are about to do.
Restoring is deliberately not a button. Swapping a database under a server that
is writing to it means stopping every writer, replacing the file and reopening,
and a half-done version of that destroys the thing it was restoring. The panel
says what to do instead.
`RunBackupNow` turns out to have been a second copy of the backup logic, written
for a button that did not exist yet, and the two had already drifted: it never
created the backups directory and logged nothing. One implementation now.
Administrators only, and sessions only. A backup is a complete copy of the
database, so this hands out everybody's data in one file.
Co-Authored-By: Claude Opus 5 <[email protected]>
Three things reported from use, all in the sidebar.
Projekter, Delt med mig, Filtre and Etiketter fold, the way a project group
already does. Kept on the account rather than in localStorage, and for the same
reason the group's own `collapsed` is: folding a heading says "this is the part
of my work I am not in right now", and that is as true on the laptop as on the
desktop. The sidebar's *width* is the opposite case and stays per browser. A JSON
array of keys rather than a column each, so a renamed heading is not a migration.
The group heading is no longer indented — only what is filed under it. The colour
dot sat as a sibling before the fold button and pushed the whole row in, and a
heading that starts further in than the rows above it says the heading is inside
something.
And the colour is chosen while renaming. The dot was itself the button, which
gave the heading four controls on a line narrow enough to crowd the name; both
are edits to the same thing, so they belong in the same moment. The form closes
on `focusout` and only when focus has left it entirely, and the swatches refuse
focus on mousedown — Safari and Firefox do not focus a button on click, so
`relatedTarget` would be null and the form would shut under the pointer.
Co-Authored-By: Claude Opus 5 <[email protected]>
A month grid is anchored to a month, so the two days on either side of its edge
sit in different grids — and you cannot page mid-drag, because a drag in flight
swallows the click that would page. A week that straddles the 31st and the 1st
has both days in the same row.
The same component as the month, with `span="week"`, rather than a second one:
the part that is easy to get almost right is the same in both. A drop target has
to decide from `dataTransfer.types` alone whether it accepts, and two copies of
that are two places to be almost right.
Seven cells instead of forty-two, so the height a month spends on six rows goes
into one and the cells hold a real list rather than a hint. On a phone the row
stacks and each day carries its own weekday name, because a column of bare
numbers is not a week.
`calendar` still means the month in localStorage. It is the value already stored
for everybody who chose it, and renaming it would silently reset them.
Co-Authored-By: Claude Opus 5 <[email protected]>
The other half of the error log. That one is what broke; this is what was done,
and nobody could ask it: `ListActivity` takes one project id, and an
administrator is not necessarily a member of the project they need to look into.
Paged with a keyset cursor rather than an offset. The log is written to
constantly and read rarely, so a page fetched a minute after the one before it
would be shifted by everything written in between — one row shown twice and the
row behind it skipped, which is the single failure an audit log must not have.
Both halves of the cursor matter: the log is written in whole seconds, so a
timestamp alone cannot separate two rows written in the same one.
Migration 0009 makes `activity.user_id` SET NULL for the same reason 0008 did it
to `created_by`. A log whose rows vanish with the account they describe answers
"what has happened here" with everything except the part somebody had a reason to
remove. `project_id` keeps cascading on purpose: the trash is the undo, and past
it the deletion was deliberate.
The event names moved to `web/src/lib/events.js`, because two surfaces read the
same log now and two copies of nineteen strings is two places for them to drift.
Co-Authored-By: Claude Opus 5 <[email protected]>
`tasks.created_by` was ON DELETE CASCADE, so retiring a colleague removed every
task they had ever written — including in projects somebody else owns. The
interface said so in plain words before you pressed the button, which is a
warning rather than a fix. It was the last remaining way to lose other people's
work by accident.
SET NULL instead: the task stays and loses its author, the same shape
`assignee_id` and `completed_by` already have.
SQLite cannot alter a foreign key action, so the table is rebuilt — and with
foreign keys on, the drop in the middle of a rebuild is a cascading delete that
takes every row referencing `tasks` with it. `PRAGMA foreign_keys` is a no-op
inside a transaction, so migrations now run on a connection of their own, and a
migration can opt out of enforcement with `-- verdande:rebuild-tables`.
`PRAGMA foreign_key_check` runs before the commit, so a rebuild that leaves a
dangling reference fails the migration rather than writing it.
The test fills a database at 0007, upgrades it, and counts. Without the marker it
loses a task, a label link and a comment — which is exactly what it is for.
The admin list now says two numbers instead of one: what disappears, and what
stays behind unattributed. They are different sentences and the second is a
reassurance.
Co-Authored-By: Claude Opus 5 <[email protected]>
The frontend build hung on 0% CPU, roughly every other run: blocked in read()
inside libuv, no output, forever. `svelte-kit sync` alone did it too, so it was
not the bundler.
The cause was not the toolchain. `~/Documents` is iCloud-managed, and iCloud
evicts what it thinks is cold — which for a dependency tree is nearly all of it:
359 of 400 sampled files under node_modules were `dataless`. Reading an evicted
file blocks until it is downloaded again, and a download that stalls hangs the
build for good. The tell was a different error entirely: Playwright once reported
`ETIMEDOUT: connection timed out, read` while loading jsconfig.json, and
ETIMEDOUT on a local file read does not exist.
Vite 5 and esbuild 0.21 predate Node 22 and this machine runs Node 26, so they
went up anyway — to Vite 8 with @sveltejs/vite-plugin-svelte 7, which SvelteKit
2.70 takes. `package-lock.json` had to be deleted for npm to resolve it; it held
plugin 4 and rejected the whole tree with ERESOLVE. That made the build faster.
It was not what was hanging.
`engines` says Node >= 22.12 now, so the next person is told rather than left
watching a process do nothing.
Co-Authored-By: Claude Opus 5 <[email protected]>
The task drawer was `width: 100vw`. A viewport unit includes the scrollbar
gutter and a percentage does not, so the drawer came out a few pixels wider
than the page it was pinned to and the whole document could be shoved
sideways. It is `100%` now, and at phone width it is stated as four insets
instead — a box defined by the viewport's own edges cannot be a rounding
error wider than it.
The scrolling body gets `min-width: 0` and `overflow-x: hidden` too, so a
pasted URL or a long unbroken word in a comment cannot stretch the flex
column and take the page with it.
Writing the test for that found another one: creating a project from the
sidebar navigated to it and left the drawer covering the screen. Every link
in there closes it on the way out; the form did not, because it calls
`goto` directly. Now it does the same thing for the same reason.
The phone test checks the document cannot be scrolled horizontally, both
with the drawer shut and with it open.
Co-Authored-By: Claude Opus 5 <[email protected]>
On a phone, a project called "Skovvænget" rendered as a vertical column of
single letters while the buttons ran off the right edge. It looked like a
broken font.
The header is a flex row of a title and four controls, and 390px does not
have room for them. Nothing wrapped, so the heading was shrunk to its
min-content — which with `overflow-wrap: anywhere` is one character. Adding
Historik this evening is what tipped it over.
It wraps now, and the title keeps a floor of 8ch so flex stops squeezing at
the point where wrapping is the better answer. Kommende had the same shape
and gets the same fix.
Every other smoke test runs at desktop width, which is exactly why this
shipped. There is one at 390 now: it measures the heading's box — wide
enough to hold a word, short enough not to be a column of letters — and
checks the page cannot be scrolled sideways.
Also stops an empty project saying both "Uden sektion" and "Ingenting her
endnu": that label is a heading for the tasks outside a section, and with no
sections there is nothing for them to be outside of.
Co-Authored-By: Claude Opus 5 <[email protected]>
Two of the last five releases died at exit 132 — SIGILL, an instruction
QEMU does not implement — partway through `npm run build`. It reads as a
broken build and is a broken emulator.
The cause is that neither builder stage was pinned to the build platform,
so buildx ran both once per target: Node and esbuild under emulation for
arm64, and the entire Go toolchain too. Nothing in either stage needs to be
the target architecture. The frontend produces HTML, CSS and JavaScript,
which are the same bytes wherever the image runs, and Go cross-compiles
natively with CGO already off.
So both are `--platform=$BUILDPLATFORM` now, and the binary is built with
`GOARCH=${TARGETARCH}`. The frontend is built once instead of twice, and
nothing is emulated at all — which should also take most of the fifteen
minutes off a release.
Co-Authored-By: Claude Opus 5 <[email protected]>
The panel's watcher reported four 5xx today — twice from one address at
11:49, twice from another at 10:45 — and there is no way to find out what
any of them were. verdande logs the real cause of every 500, but on a Rune
the log belongs to the container and the container is replaced on every
restart. So an operator is told that something broke and given nothing to
act on, which is the worst of both: alarming and useless.
Every 500 now writes a row: what the handler was attempting in its own
words, the error it hit, the method and path, who was asking, and the
request id that ties it back to the log line for whoever still has it. The
"what" is the part a status code cannot give — "list projects" and "move
task" are the difference between a fault you can find and one you can only
count.
Written best-effort and never checked. This is the path that is already
reporting a failure, and a diagnostic that can itself fail the request
would turn one broken screen into two.
Threading the request through `internal()` and `storeError()` touched 108
call sites across 19 files, which is the price of not having had it there
from the start. It is mechanical and the compiler checked all of it.
Kept for thirty days by the nightly sweep: long enough to answer "what
happened last week", short enough that a fault loop cannot fill the volume
the tasks live on.
Read at Indstillinger → Fejl, administrators only — it carries paths,
account names and the server's own error strings. The empty state says
"Ingen fejl registreret. Det er den rigtige tilstand", because a blank
panel should not read as a broken one.
The test drops a table and asks for the endpoint that reads it: the same
code path a disk error takes, and the row that comes back has to name the
operation, the account and the request.
Co-Authored-By: Claude Opus 5 <[email protected]>
The manifest pointed at icon.svg and nothing else. Chrome and Firefox
install happily on that; Safari does not read the manifest's icons at all
and looks for /apple-touch-icon.png by name, so an installed PWA on iOS got
a screenshot of the page as its home-screen icon.
This was written down as blocked: there is no SVG renderer on this machine,
which is why the references were removed rather than left pointing at files
that did not exist. Playwright is an SVG renderer, and it has been a
dependency the whole time.
scripts/generate-icons.js renders the same SVG at 192, 512 and 180. A script
rather than a test, so CI never runs it and never rewrites a tracked binary
— run it when the icon changes. The smoke test already fetches every icon
the manifest names, and now asks for the Apple one by name too, so a
reference with no file behind it fails there rather than on somebody's
phone.
Co-Authored-By: Claude Opus 5 <[email protected]>
"Do it today" is the most-made rescheduling there is, and the sidebar is
where the pointer already is when you have picked a task up.
Only that one of the two date views. "Kommende" is a range rather than a
day — it starts today and runs a week — so a drop on it would have to
invent a date its own label does not name. The month grid inside it takes a
drop per cell, which is where a specific day is actually on offer.
Co-Authored-By: Claude Opus 5 <[email protected]>
Two more things the database has been recording and the interface has
never shown.
**Who a task is waiting on, on the row.** Only when it is not you: your own
list would otherwise carry your own face on every line, which is a column
of one initial telling you nothing. The point of showing an assignee is
that it is somebody else.
That needed names, and a task carries only an `assignee_id`. `GET /people`
is the primitive: everybody you share a project with, plus yourself,
fetched once per session. Not per row, and not per project either — a list
would have to know which project each task belongs to before it could ask
for that project's members. Deliberately *not* the instance's user list,
which is `/users` and administrators only: a task list has no business
enumerating everybody with an account here.
It sits first in the meta line, because who changes what the rest of the
line means — a date on somebody else's task is their deadline, not yours.
**And the activity log.** Nineteen kinds of entry have been written since
the beginning — every project, section, member, task and comment event —
and `GET /projects/{id}/activity` has answered honestly the whole time with
nothing on the other end. There is a Historik panel on the project now,
loaded when it is opened rather than with the page, and the event keys
become Danish in exactly one place. An unknown key falls through to itself
rather than being hidden: a build that records something this table has not
learned should still show that it happened.
One test assertion had to change, and it is worth saying why. It matched
"males inden fredag" against a row whose title is "males inden" — quick add
read "fredag" as the due date — so it was really matching the title
concatenated with the metadata beside it. Putting the assignee between them
broke it. The assertion now names the task's own text.
Co-Authored-By: Claude Opus 5 <[email protected]>
An invite sent with the wrong role could not be corrected. The only way
back was to remove the person and invite them again — and removing a member
unassigns every task they were responsible for, so fixing a dropdown cost
somebody their work. This is the other side of "an external user cannot
create tasks": the rule was right, the role was wrong, and there was no way
to say so.
PATCH against a membership that already exists, not the upsert AddMember
does. Somebody who is not a member is not silently made one: that is an
invite, it belongs to the invite flow, and doing it here would let a typo
in a user id add a stranger to a project without anybody being asked.
The owner is refused, with the real reason rather than the true but useless
one. Ownership lives on the project rather than in project_members, so the
plain answer would have been "not a member" — which is correct and tells
nobody anything. Ownership is transferred, not granted, and two owners is a
state with no way to settle a disagreement about who may remove whom.
In the list, the owner's standing is text and everybody else's is a
dropdown. A disabled control would be a thing that exists to say no.
And the roles are finally in Danish. The API's words are `owner`, `editor`
and `viewer`, and they were being printed straight into a Danish page; the
list now uses the same three phrases the invite form beside it already did.
The smoke test drives both browsers, because the interesting part is that
the other one's permissions really change: a viewer is told the project is
read-only and has no quick-add box, the owner promotes them from the list,
and after a reload the box is there and the notice is gone.
Co-Authored-By: Claude Opus 5 <[email protected]>
Two things from the same report — an external member who could see a
project but could not get work into it.
**A section with no tasks could not be filled by dragging.** Only the task
rows accepted a drop, so the target was always another row; a section with
none had nothing to aim at, and a section you have just made is empty by
definition, which is exactly when you want to put something in it. The
board has taken a drop on the column since it was written. The list now
does the same: the section is the target, and a drop on it puts the task at
the end. Dropping on a row still means "here exactly" and stops there
rather than bubbling to the section behind it.
The unsectioned block at the top is a target too, which is how a task comes
back *out* of a section — previously only possible if some other loose task
happened to be there to aim at.
**And the page said "Projektet findes ikke, eller du har ikke adgang til
det" for three different things.** `project` started null with no loading
state, so that sentence was what you saw before the first request returned,
what you saw when it failed, and what you saw when you genuinely had no
access — and the effect never asked again, so the wrong one stayed until
somebody reloaded. That is very likely what was reported: the container was
restarted twice while people had the app open, and a request that lands
mid-restart leaves the page insisting the project does not exist.
Four states now. Loading renders nothing, as the shell does. 404 keeps the
old sentence, since the server deliberately answers 404 for both "no such
project" and "not yours" and the interface must not guess between them.
Anything else says the request failed and offers to try again.
What is *not* a bug: a viewer cannot create or move tasks. Checked at the
API — an editor invited by link gets 201 on create, 201 on quick add and
200 on a move into a section; a viewer gets 404 on all of them, which is
the rule working. If the person reporting this cannot add tasks, their
membership is viewer rather than editor.
Co-Authored-By: Claude Opus 5 <[email protected]>
The invite test proved the shared project reached the invitee's sidebar and
stopped there. A member's project page is a different path from an owner's
— different role, different guards — and it was never opened by anything.
It also matters that the page shows "Projektet findes ikke, eller du har
ikke adgang til det" while it is still loading, because `project` starts
null and there is no loading state. A failed load and a slow one look
identical, which makes this exactly the kind of thing a person reports and
a test suite cannot see.
Signs in as the invited account in its own context and clicks through to
the project, asserting the heading renders and that message is absent.
Co-Authored-By: Claude Opus 5 <[email protected]>
`projects.color` has been in the schema since the first migration, has
defaulted to 'graphite' on every row ever written, and nothing has ever
read it. There was no palette, no name-to-colour mapping, and the dot in
the sidebar was hardcoded to a border grey. So this is not "add a picker";
it is implementing colour for the first time.
Names, not hex values. A hex in ten thousand rows is a decision nobody can
revisit; a name is resolved when the interface paints and can be retuned
for a new theme or a new display. It is also the difference between a
database that stores a fact and one that stores a rendering. The set is
closed and refused at the edge — an unknown name would be stored happily
and painted as the default, which looks like a colour that did not save.
One palette across all five themes, and this is the one place in app.css
that is not redefined per theme. Everything else there is text, a ground or
a border and has to be measured against what sits behind it. These are
small solid marks beside a label that already carries the meaning: what
they owe is to be told apart, and to be visible on both a near-black and a
white ground, which is why all ten sit in the mid-tone band. It also keeps
them out of the trap the themes are in — a colour added to four blocks and
forgotten in the fifth looks finished and falls back silently. There is no
fifth block to forget.
The list is written down three times: Go for validation, colors.js for the
picker, app.css for the tokens that paint. So a test reads the other two —
the same small-scanner approach openapi_test.go takes to the spec — and
fails if the server accepts a colour nothing can draw, or the picker offers
one the server refuses.
Groups get a colour of their own (migration 0006) and projects inside a
group are indented. The indent is to the group's *name*, not to its dot:
the eye follows the text, and lining the projects up under the dot leaves
them looking a step short of where the heading starts. The drop line is
indented with them, so a drop inside a group reads as landing in the group.
The group's dot is the control as well as the preview — you press the thing
you are about to change, as on the project page. That was not the first
design: a third text action, "Farve", sat beside Omdøb and Slet, and those
carry `opacity: 0` rather than `display: none`, so they take their width
whether or not you can see them. A group called "Arbejde" rendered as
"ARBEJD". Moving the control onto the dot gave the name its room back.
The smoke test measures the indent at the content edge rather than the
element's, because the indent is padding and the box does not move — which
is how the first version of that assertion passed 12 against 12.
Co-Authored-By: Claude Opus 5 <[email protected]>
The README has promised "internal users are created by an admin" since the
beginning and there was no such thing. There are exactly two ways an
account has ever come into existence: the first admin at setup, and
somebody accepting an invite to a project. So the only way to add a person
was to share a project with them first.
The store has been ready the whole time. CreateInvite already documents
that "projectID may be empty, which invites somebody to the instance
rather than to a particular project", and AcceptInvite already has the
branch that grants no membership when there is none. requireAdmin has
existed in the middleware and was used nowhere. This wires those three
together and puts a page on them.
An invite rather than an account with a password the administrator picks.
A password typed here would be known to two people before it was ever
used, and would then have to travel to the second one by some channel
nobody has thought about. No name is asked for either — the signup form
asks the person, and a name an administrator guesses is a name somebody
has to correct.
Two refusals keep the instance administrable: the last administrator
cannot be demoted or deleted, and nobody can delete the account they are
signed in as. Both are enforced on the server; the interface simply does
not draw the buttons on your own row, because a button whose only purpose
is to be refused wastes a click to teach a rule. The last-admin refusal
gets its own error code, because "that already exists" is not what
happened and this is the one refusal a person needs explained.
Deleting an account is the only hard delete in this database, and building
the confirmation is what turned up how wide it goes. `projects.owner_id`
cascades, which is expected. But `tasks.created_by` cascades too — so
removing a colleague deletes every task they ever wrote, including in
projects somebody else owns. The count in the list says so, and the
confirmation says it in words. Changing that to ON DELETE SET NULL, so the
work stays and loses its author, means rebuilding the tasks table and its
FTS triggers: worth doing, not worth doing quietly while adding a screen.
It is written down in the store instead.
Pending invites are listed and can be withdrawn. Only the token's hash is
stored, so a link sent to the wrong address cannot be looked up and
resent — withdrawing it is the only way to stop it before it expires.
Six API tests for the guards and the counts, and two smoke tests: the
whole loop from invite to a second browser creating the account and
appearing in the list as somebody who is not an administrator, and the
absence of buttons on your own row.
Co-Authored-By: Claude Opus 5 <[email protected]>
