Repository navigation
v0.1.5
Added
A dark theme, and a light one, either chosen or left to the machine.
The default is to follow the machine, and that is a standing instruction rather
than a value somebody was assigned: it keeps following, so a screen that goes dark
in the evening takes the app with it. Nothing detects the setting once and stores
the answer, which sounds like the same thing and is not. A stored answer cannot be
told apart from a decision, so the app could either follow the machine or remember
a choice, never both — and the mechanism that would have written it only fires for
an account that has never chosen anything, which no existing account qualifies as,
because saving a timezone counts. Every one of them would have upgraded into a
light app on a dark machine.
So there are three states, not two. Light and Dark stay where you put them.
Follow my system is the default and is the media query, which means it is right on
the first paint with no JavaScript at all. The choice is on the account rather than
in the browser, so it travels to another device, and a theme column defaulting to
system lands every existing account on the honest answer with no backfill to
guess at.
The moon in the sidebar switches it, next to Sign out — the only part of the shell
that already holds controls about this session rather than about the ledger. It is
a plain button that says what it will do, not a switch reporting what is on: a
two-state control cannot honestly report a three-valued setting. Pressing it while
the setting is Follow my system resolves the machine and sets the opposite
explicitly. Settings has the three-way choice, applying as you pick it rather than
on a Save button, because it is the one preference whose result is visible while
you are choosing it and two controls for one value must not be able to disagree.
Painting it before the page paints took a file rather than the usual inline script,
because the Content-Security-Policy here is script-src 'self' with no nonce and
no hash. An inline script would have been refused, reported nowhere — nothing
declares a report endpoint — and would have looked perfect locally, since neither
dev server applies the policy and one of them never serves the shell at all. It
would have reached a release as a flash of the wrong theme on every load.
The stylesheet had one palette written into it in 189 places. It now has two, in
one place each: 57 tokens, every one declared in both themes, with a test that
fails on a colour written anywhere else and on a token given a value in only one
theme — which is the bug that makes half an app unreadable while looking fine to
whoever wrote it. Four literals turned out to be two colours sharing a spelling:
white is both a card and the text on a green button, and only one of those is
still white in the dark. The same split runs through the accent, where the green
that reads as a link is not the green a button is filled with, and in dark the
first has to lift while the second stays dark enough to carry white.
Three repairs to the light theme came out of writing the second one down. Six
greys carrying real text were under the contrast a person needs, the input
placeholder worst at 2.65:1. An input's border was 1.39:1 against the field it
edges, which is not a boundary — and an input here is white on a white card, so
that border is the only thing saying where the field is. It holds 3:1 now, and a
focused field is darker again rather than only greener, which is what it had
briefly become once the resting edge moved up to meet it. And the focus ring took
its contrast from whatever happened to be behind it, because it was
semi-transparent; it is opaque now.
The report palette was worse than it looked. Under simulated deuteranopia its
green and its pink were 1.78 apart as CIEDE2000 measures it, which is to say they
were the same colour, and that shipped in 0.1.4 when six colours became ten. The
two palettes now reach 5.6 and 4.7 by keeping each slot's hue family across both
themes and varying lightness, which is the channel that survives. An honest limit
on that: ten categorical colours cannot all be told apart by somebody with
dichromatic vision, and a search that held hue identity and the contrast a line
needs could not beat about 7 and 4. The remedy is a second channel that is not
colour, which is a change to the charts rather than to the palette. Until then the
legend and the table under every chart carry identity, and both are always there.
Emailed notifications, on a schedule, in two kinds.
A recurrence can say to write when it proposes. One message per proposal however
many rows it holds, naming the dates and pointing at Staged transactions, sent to
the address on the account. It says nothing on a tick that proposes nothing, so a
schedule that has caught up goes quiet rather than arriving every five minutes.
The Recurring list says which recurrences are set to send one.
A template can carry a reminder, which is the other half of the same idea: a
template is filled in by hand, so nothing can make it for you, but something can
tell you it is the day. Once on a date, or repeating on the same schedules a
recurrence offers, and either way at a time of day — the first thing here that
needs more than a date, and it is read on the person's own clock rather than the
server's. The Templates list says which templates have one and whether it
repeats.
A reminder that happens once is a first-class answer rather than a yearly rule
nobody means: its frequency is null, and it refuses the interval and the policies
a repeating rule needs rather than storing leftovers of them. Once sent it says
so and the scheduler stops looking at it.
Both ride the recurrence scheduler's existing tick rather than a loop of their
own, and a backlog collapses into one message: coming back from a week of
downtime brings one reminder, not seven. Neither can be delivered without
SMTP_HOST and MAIL_FROM, and the form says so rather than accepting a setting
that would quietly never fire.
Reports. Six of them, over one date range, per currency and never added across
currencies: net worth and a balance sheet for what the accounts hold, income
against expense and categories for what moved, a cash flow statement for where
the money that can be spent came from and went to, and a trial balance that
totals zero when the books are whole. Each is a preset over one query that
differs by which accounts it reads, whether it reports a period's movement or
the balance it ends on, and how time is bucketed — weekly, monthly, quarterly,
yearly, or not at all. Reachable at /reports and over MCP as get_report.
The category report covers income as well as expense. Spending by category on
the dashboard answers only where money went, and the same question about money
arriving had no answer.
A per-account register: every posting in date order with the balance before and
after it, and the balance the window opens and closes on. It is for finding
mistakes rather than for analysis — where a balance goes wrong, this is the row
it went wrong on. An archived account ends at zero with the postings that closed
it out to equity still in the list. On the account page behind Show register,
fetched only when asked for so the ordinary visit costs nothing, and
get_account_register over MCP.
whoami says whether this deployment can send mail at all. A reminder is stored
whether or not it can, so an agent could set one up and had no way to tell
somebody it would never arrive.
The Reminder and Notifies columns sort, ranked by what they are going to tell
you rather than alphabetically by the badge text: a reminder still to come above
one already sent above none at all.
The reminder section of the template form is laid out like the schedule section
of the recurrence form, which is the same kind of thing and did not look like it.
It was a bare fieldset, so it took the browser's own padding-inline and its
heading sat indented from every other label on the screen, with eleven controls
stacked against each other for want of a gap. It is the same card now, with the
same small-caps heading, and it ends the way schedule ends: with the next five
dates it will actually send on, and the time each goes at. The note explaining
what a reminder does moved up under the checkbox it explains, from the bottom
where it read as a footnote to the weekend policy.
The notifications section of the recurrence form had the same bare fieldset, one
section below the schedule it was supposed to match. And the template form's
description field was half a row wide, alone inside a two-column grid.
Axes on the report charts, which shipped without any. A line that ended higher
than it started was all either chart actually said: there was no value scale
beside it and no date under it, only a caption naming the range. Both now carry
gridlines on round numbers — worked out in scaled integers, so a tick sits
exactly where its own label says it does — and the dates of the periods they
cover, written as that period is named rather than as a full date, so a year of
months fits.
The axis text is HTML rather than drawn into the SVG. The drawing scales to its
panel, so a label inside it would read as twelve pixels wide on a desktop and
four on a phone, which is an axis nobody can read on the device most likely to
need one. How many dates fit comes from measuring the chart rather than from a
breakpoint, because the same chart is wide on the reports page and narrow in a
card at the same viewport.
Reports sits after Recurring in the sidebar rather than after Transactions. The
sidebar reads in order: where the money is and what moved it, then the work
waiting on you, then the things that file and repeat it, then what it all adds up
to. A test holds the order, because an ordering nothing asserts is one a later
edit reorders by accident.
Each account on the overview opens that account. The whole row rather than the
name, because the balance is what somebody is looking at when they decide to go
in, and the date range travels with them, so the account page opens on the
period they were reading.
The cash flow statement will not agree with income and expense, and the gap is
widest for whoever uses a credit card most: a purchase is an expense the day the
card is swiped, while the cash leaves when the bill is paid, in a different
period and as borrowing rather than spending. Both figures are right. The report
says so on the page rather than leaving it to be discovered.
Staged transactions finds a row that repeats something already recorded,
not just one that repeats another row still waiting. The check it had wanted the
same day and the same payee, which a real import has neither of: the bank posts
when it settles and names the merchant its own way. So the amount is the anchor
now, with the account and the direction to keep two unrelated spends of the same
size apart, and three days of latitude on the date. Payee and category are
ignored, being the two most likely to differ between a bank's record of a
purchase and yours. The same duplicate filter finds all of it.
That looser test is advisory. What refuses a commit is unchanged and still
strict, because loosening it would start turning down two genuine coffees bought
on one card in one week.
A run through every flagged row, from Review N possible duplicates at the
top of Staged transactions. Each comparison says which one of how many it is and
carries Previous and Next, and dropping a copy lands on the next pair rather than
back at the list, so a dozen duplicates are a dozen decisions instead of a dozen
round trips through the queue to find the next badge. The badge on a single row
still opens that pair directly.
A side-by-side review, reached from the badge on any flagged row. Both records
are open to edit and each saves on its own; a staged row saves rather than
commits. The one already in the books sits second — on the right, or underneath
on a phone — and where both are staged the older one does. Only a staged side
can be dropped, because the way out of a duplicate is to remove the copy that
has not been recorded yet. It is not a diff: the fields that differ are the ones
that always differ, and colouring them says nothing a person reading two
transactions cannot already see.
Changed
How this describes itself. Every page inside the app already spoke plainly —
"Where your money sits and how it moved", "Everything you track, from checking
and cards to cash and crypto wallets", "Rows waiting on you" — and the marketing
copy was the one place that talked like a spec sheet. It led with the automation,
then with double-entry, then with privacy, none of which is what somebody wants
from a personal accounting app; they are how it delivers what somebody wants.
So the front page, the manifest, the container label and the sign-in screen all
lead with what you get: where your money is and where it went, every account in
one place, statements that file themselves, bills you set up once, reports that
add up. The double-entry books and the MCP server follow as the reasons those
things can be relied on — which is the honest ordering, since the books are what
the agents are being careful with. The Helm chart's own description and the
scheduler image's label had been left on the old framing and now match: those are
what helm search and a registry listing show, so they are the front page for
anybody who arrives that way.
The queue in front of the books is called Staged transactions, everywhere. Twenty
places in the app called it "the review queue" instead, which is what it is and
has never been what it is called — so a recurrence promising to propose "into the
review queue" was naming a screen that is not in the sidebar under that name, and
a person who went looking for it would not find it. The name is used wherever
somebody has to go and find it, including the notification mail's subject line;
the description is kept where the sentence is explaining what the screen is for.
One verb phrase covers it throughout — a recurrence adds a row to Staged
transactions — rather than five ways of saying the same thing. The two dialogs on
the page itself say "the queue", because you are already standing on it, and the
consent screen describes what the agent can do rather than where the rows land,
which is what the three scopes beside it do.
An idempotency key means the same thing over MCP as it does over the HTTP API.
Ten MCP writes kept a replay record of their own on top of the one the service
they call already keeps, and the two matched a retry differently: the outer one
against the request as it arrived, the inner against what the service had
normalised. A retry of a mass edit that listed the same rows in a different
order was a different request to one and the same request to the other. They
call their service directly now. Records already written are inert, and no key
in an existing database loses its replay.
set_preferences no longer advertises a call it refuses. Every field of the
patch is optional, so an agent reading the schema was told that sending nothing
but an idempotency key was valid, and found out otherwise at runtime.
create_transaction says that a near-identical entry is refused and what
allowDuplicate is for. Both mass edits say that one split anywhere in a
selection refuses a category or type change for the whole call.
list_payees described itself as returning canonical payee names. It returns
every stored spelling, one row each, which is the opposite and the reason
list_duplicate_payees exists. Both now say which question they answer.
Recurrences report their shape over MCP. get_recurrence and
list_recurrences declared it as an unknown value, so the one thing an agent
reads a recurrence for was the one thing the tools would not describe.
Recategorising the last transaction off a category removes that category. Only
what an edit moved off is considered, so one made ahead of time and standing
empty on purpose is left alone, and anything a recurrence or a template still
names is kept — neither holds a foreign key, so nothing else would stop the
delete and what would be left is a standing instruction naming a category that
is gone. A queue-scoped agent edits the row and leaves the category, on the same
rule that stops it creating one.
Payees needed no such change and got none. Every list of them is a group-by over
the rows that name them, so a payee nothing references has already stopped
existing.
Fixed
The keyboard could not work the choices these forms offer. Six sets of radio
buttons had no shared name, which is the attribute that makes a set of radios
one group rather than several unrelated controls — so the arrow keys did nothing
and every option was its own tab stop, putting four presses between a four-option
group and the field after it. React enforces one-of-many through its own state
regardless, which is why it survived: it looked right, it clicked right, and only
the keyboard was wrong. The names are per instance now, so two of the same form on
one page do not share a group and clear each other.
The three transaction type grids are one control, and it now does what it always
claimed. All three declared themselves radio groups with radio children and
implemented none of what that promises, so a screen reader was told to expect a
keyboard interface that was not there. The group is one tab stop, the arrows move
the choice and wrap, and Home and End go to the ends. One of the three was not a
radio group at all: on a template, clicking the chosen type again is how somebody
says there is no type, and a radio has no way to become unset — that one says what
it is, a set of toggles.
Filtering Staged transactions by a payee could return nothing at all. The
comparison folds Unicode presentation forms on the way in — that is what makes
"fi" and "fi" the same payee everywhere else — and this one place did not, so a
payee holding a ligature or a full-width letter never matched anything and the
queue came back empty rather than saying why. The rule is now spelled the same
way in all six places it appears in SQL, and a test fails when they diverge,
because a spelling that differs does not raise: it silently fails to match.
Merging payees left a staged row's fingerprint describing the payee it used to
have. The queue flags a row as repeating another by comparing those fingerprints,
so two rows that had just been made identical stopped being reported as repeats
of each other. Rows keyed on a bank's own reference were never affected, which is
why this was easy to miss.
Resolving a payee to the spelling the ledger already keeps stopped reading the
whole ledger. It runs on every single transaction write, and the expression it
matched on — the same Unicode folding as above — was one no index could serve, so
each save scanned the account's own rows. Both sides of it are indexed now: on
five thousand transactions that is 6.3ms and two hundred buffers becoming 0.02ms
and three.
Staged transactions died on a single row it exists to show. A draft date matching
the shape of a date but naming no real day — 2026-02-30, 2026-13-01 — was
accepted, filed with a "Date is not valid" issue, and from then on every attempt
to list the queue failed on the cast. The queue was the only place that row could
be seen, so the only cure was deleting it by an id nothing would show you. The
guard now asks what validation already decided instead of trying to out-guess the
calendar. The amount guard had the same hole with no bound on length, where a
draft of two hundred thousand digits overflowed numeric and took the default
sort down with it; both now hold to the shape the domain accepts.
Two answers to what day it is. PostgreSQL reads a bare offset timezone with the
POSIX sign convention and Intl reads it as ISO, so for anyone whose stored
timezone was an offset rather than a zone name the two disagreed by sixteen
hours. An account archived in that window closed on a day the dashboard had not
reached, and its balance was left out of the headline total while the ledger went
on counting it. One implementation now answers the question, and PostgreSQL is
not asked.
Net worth and the balance sheet took an archived account's history away along
with the account. Excluding it dropped it from every bucket, including the months
it was open and holding money, so a monthly chart lost history it had reported
correctly the day before. Archiving posts the balance out to equity, which is
what carries the account to zero on the day it closed; the row is kept and hidden
only when it is flat at zero across the whole window. So on those two reports
includeArchived no longer changes a figure, only which rows are listed, and
each row now carries archived — a closed account's past would otherwise read as
money still sitting in the live ones. A currency with nothing left to list is
left out rather than returned as an empty section.
A payee-sorted list paged with a cursor could skip rows or never end. The
ordering was lower() in the database and toLowerCase() in JavaScript, and
those are different functions — they disagree on a Turkish dotted capital and on
a final sigma, and whether they disagree at all depends on the database's
collation. The cursor now carries the value the database sorted on.
The categories report added income to expenses and headed the answer "Net". Each
row is a magnitude there, on purpose, so the column sum is a total filed rather
than a net, and it says so. The trailing column of a running-balance report is
headed "Closing" rather than "Total", because that is what it holds.
The same report showed one name twice with no way to tell the rows apart. A
category set to cover both sides is two rows there, correctly — one for what came
in under it, one for what went out — and Uncategorized always is. Only a name
that really does span both sides is qualified now; every other row reads as the
person wrote it.
A report chart had six colours and no limit on how many rows it would draw, so a
seventh account shared the first account's line and the legend said two things at
once. There are ten, and a test fails if the stylesheet and the code disagree
about how many.
Every badge in a transaction row sat on a line of its own, doubling the row's
height. The class holding the payee beside it had never been defined in the
stylesheet. Six other class names in the browser named nothing at all and are
gone.
The audit log recorded a template's reminder as having never existed. Three of
the five paths that write a snapshot took a default of null, and in an
append-only record null does not read as "nobody asked" — it reads as "there was
no reminder". A bulk edit therefore claimed to have removed one, and both delete
paths recorded the reminder as never having been there.
A template saved again re-sent a reminder that had already gone. The row is
replaced whole, so an edit to the payee replaced the watermark with it and the
reminder was owed again. A schedule that really did change still starts afresh,
which is what somebody moving the date is asking for.
The rows waiting above a committed list were not narrowed by the filters the
list was. On a category, payee or template page the two read as one list, so
choosing a type or typing into the search box moved the committed rows and left
the staged ones standing above a list they had just been excluded from.
The register's link on each posting carried a parameter no list in the product
accepts, so it was dropped and the row landed on the whole unfiltered
transactions list. It narrows by account and that one day now.
The application shell described the product differently from the manifest the
release publishes, and coloured a phone's browser chrome near-black on a
stylesheet that commits to a light scheme.
A staged mass edit left behind a category the identical edit, done one row at a
time, would have cleared. The committed side had the same split and was fixed in
this cycle; the queue had it too. A caller holding only ledger:stage still
leaves the category standing, on the rule that a queue token proposes and never
decides.
The reminder sweep claimed rows on a deployment with no mail server. Every claim
advanced a watermark past an occurrence nobody was told about, so configuring
SMTP a month later would have found a schedule that had quietly eaten its own
backlog — the opposite of what the form promises. It leaves before claiming
anything now.
The split deployment served the one document that runs the app without seven of
the headers every other response on the origin carries, including both
Cross-Origin-* policies. nginx repeats the whole set now, the list lives in one
place, and a test runs a real response through the middleware and fails if the
two differ by a header or a value. X-Frame-Options is DENY on both, rather
than Hono's SAMEORIGIN contradicting the frame-ancestors 'none' beside it.
The generated first-run setup code belonged to one process. On a web tier running
more than one replica — which the chart does by default — the code printed in the
log was rejected by every other pod, so the claim the chart's own notes describe
failed about half the time. It is stored now, like the MCP signing key, so every
replica agrees on it. An operator-chosen SETUP_TOKEN still never touches the
database, and Pulumi can supply one, which it could not before.
Validating a staged draft opened a ledger account. Preparation is meant to answer
whether a draft would balance and nothing else, so an agent holding only
ledger:stage — the posture that is supposed to be unable to touch the books —
added a counter-account and a new zero row to the trial balance. It looks the
account up now and stands one in when there is none.
Smaller ones. A malformed request body answered 500 with a stack trace instead of
400. A mistyped /api/v1 path, and any with a trailing slash, came back as the
application shell with a 200. Responses carrying a session token had no
Cache-Control. A broken consent cookie 500ed. An APP_BASE_URL that was not a
URL, and every strict scalar setting, refused to start without saying which
variable was wrong. A register window opening after today summed future postings
into a balance labelled as of today. The cash flow statement read every posting
in the ledger to answer about one month, at eight times the cost. A failed report
also told a full ledger it was empty, and reports could show figures from before
an edit. The recurring list ordered amounts as text, so 1.50 sorted below 1.45.
One search box had no accessible name.
A table wider than the panel holding it spilled past the edge instead of
scrolling inside it, so the far columns could not be reached at all. The
templates and recurrences pages both wrapped their table in table-wrap, a
class nothing in the stylesheet ever defined, and had done since they shipped.
The rule exists now, and a test names any table left without a wrapper that
scrolls.
Cash flow, income and expense, and spending by category asked whether each entry
still runs through an archived account once per posting rather than once per
query. On a ledger of a hundred thousand postings the dashboard's own cash flow
ran that subquery twenty-eight thousand times, reading a hundred and seventy
thousand buffers to produce two rows. The same rule is now one aggregate the
planner turns into a hash anti-join: sixty times fewer buffers and seven times
faster, with the netting that decides membership unchanged.
A duplicate payee group offers the spelling the ledger would itself keep. The
group was ordered by how often each spelling is used and then by name, while a
write reusing a payee breaks a tie by preferring a name already equal to its own
cleaned form. So three equally used spellings of one shop offered
" ACME MARKET " as the one to merge into, where the ledger would have kept
"Acme Market". Both the browser and the MCP guide say the first entry is the
target, so this was the wrong answer rather than a cosmetic ordering. One rule
now, used by the group and by the write.
Sorting Staged transactions by amount could fail on a row whose amount a CSV left
unreadable. The guard meant to catch that admitted anything with a digit either
side of any character, so 42x50 passed it and then raised on the cast — a
backslash lost to a template literal, in a regex that has been there since the
column was sortable.
The recurrence form's preview test asserted a date that has now gone by, and
would have failed from here on whatever anybody changed.
Internal
Gates that could not fail. The PostgreSQL job would have gone green having
skipped every integration file if it ever lost its database URL, and nothing
asserted otherwise. The chart's negative gate checked that a render failed
without checking why, so deleting the guard under test still read as "refused".
The NetworkPolicy template was never rendered at all, being off at default
values. A release resolved its tag twice, so a tag moved mid-run could verify one
commit and publish another; it resolves to a commit once now. A dispatched
re-publish inferred that an unsuffixed version was final and moved latest on to
a release marked as a prerelease — it asks instead. Two tags could publish at
once and race for latest. And no test covered the chart's appVersion, the one
version location whose drift installs the previous release's images.
A refused second Ralph run deleted the lock the first one was holding, because
the lock was named before it was taken, leaving a third free to start alongside
the first. The git guard's config scan missed git's one-line
[section] key = value form, catching the same key only when written underneath.
set-version did not know about the Pulumi project's manifests. They happen to
be current, because the version they were added at is still the version — but
nothing in the root install or the root verify reads them, so nothing would have
said otherwise.
Thirteen schemas in the shared contracts no longer carry an export nothing
outside the file used. Six types the browser had hand-written copies of are
re-exported from the contracts they duplicated.
CI starts the frontend image rather than only building it. Its nginx
configuration is a template the entrypoint renders at startup, so a malformed
add_header — or a location-scoped directive in an include the main config
pulls into http{} — is a startup error a build never sees, and it would have
reached a release. The shell and a hashed asset are both asked for, because
nginx drops inherited add_header from any level with one of its own.
The two lockfiles are held to the same resolved versions rather than only the
same ranges, so installing in one and not the other cannot ship an image running
a version the suite never saw.
clockTimeIn had no test. It is what decides whether a reminder goes out at the
hour somebody asked for, by string comparison against a stored HH:MM. Covered
now, along with a bare offset read as ISO rather than POSIX, both ends of the
date line, both daylight-saving boundaries, and an unreadable timezone falling
back to UTC rather than throwing inside a loop that serves every tenant.
Both .env.example files and the chart's values said a mail server buys a
password reset and address confirmation. It also decides whether any scheduled
reminder is ever delivered, which was in none of them.
set-version now rewrites the example image tags in the split-deployment
compose file and the Pulumi README. Six tags a release would have left pinned to
the previous version, on the two pages somebody copies from.
Four indexes are dropped whose leading column another unique constraint on the
same table already leads with. No query loses a plan; four tables stop
maintaining a second copy of their own first column.
Money arithmetic moved out of components.tsx into money.ts. It was three
hundred lines of exact decimal comparison filed under a name that says React
component, which is not where anybody looks for arithmetic. Two report suites
compared money by casting it to a double, which is the mistake the house rule
exists to forbid, and two brute-force property tests had unbounded scan loops:
an implementation that stopped advancing would have hung the file rather than
failing it.
The README is half the length. Everything it said about how each feature works
moved to a guide of its own, so the front page answers what this is, what is
good about it, and how to start, and the walkthrough is one link away rather
than a hundred and fifty lines down. The architecture notes now cover the mail
half of the scheduler, which had shipped undocumented.