money/CLAUDE.md

# money

Statement-driven personal finance tracker. A data directory holds one folder per
account; statements dropped into those folders are parsed into a rebuildable
SQLite index, categorised by ordered glob rules, and browsed in a web app
(`money serve`, which is also the default command).

Read `README.md` first — it is the user-facing reference for every config key.
This file covers what the code assumes and why.

## Layout

```
cmd/money/main.go     subcommands; serve is the default
internal/config       rules.toml (rules + transfers), account.toml, XDG config, data-root resolution
internal/model        Account, Transaction, amount formatting, description normalisation
internal/glob         the `*` / `?` matcher used by rules (two-pointer, no exponential blowup)
internal/parser       Parser interface + registry; nlb, revolut, traderepublic
internal/store        SQLite index (modernc.org/sqlite, no cgo)
internal/importer     directory walk, dedupe, balance checks
internal/rules        applies ordered rules, writing rule_tag
internal/transfers    pairs the two legs of a movement, writing the transfers table
internal/report       per-tag aggregation
internal/web          `money serve`: JSON API + embedded single page (static/)
```

## Invariants

**Every tag comes from rules.toml.** There is no way to tag a transaction by
hand; `rule_tag` is derived state that `Engine.Retag` rewrites wholesale, which
is what makes `money retag` safe to run at any time. Nothing may write a tag
from anywhere else — the moment something does, the index stops being
reproducible and retag stops being safe. Covered by `TestRetagRewritesEveryTag`.

**A transfer is a pair or it is nothing.** A `[[transfer]]` block names both
legs (`from_account` + `from_desc`, `to_account` + `to_desc`), and only a
matched pair is dropped from the report — both legs together, never one. This
is the whole reason the earlier boolean `transfer` flag was removed and this
replaced it: a one-sided verdict let half a movement vanish and left the report
unbalanced. An unmatched leg therefore keeps counting, and is surfaced as a
warning instead: `⚠` on the transfers screen, on stderr from `import`/`retag`.
Nothing may start excluding a single leg on the strength of one side matching.

Legs pair within `transfers.WindowDays`, nearest date first. **Within one
currency the amount is the evidence and must be the exact opposite**, unless
the definition sets `tolerance_pct` — a per-definition opt-in for a route where
the bank takes a fee on the way, so the two statements genuinely disagree. It
defaults to zero and belongs on the one definition that charges: a global or
default tolerance would loosen every route that does not. **Across currencies
the amount is not checked at all**: there are no exchange rates here, so the two
numbers are unrelated and the dates carry the pairing alone. A tolerance means
nothing there and is ignored. That asymmetry is the design, not an oversight —
do not "fix" the cross-currency case by inventing a rate, and do not turn the
tolerance into the default. Like rules, the first definition claims a leg and a
transaction belongs to at most one transfer.

**A tolerated mismatch is a fee, and a fee is money, so it is reported rather
than forgiven.** That is the condition on the tolerance existing at all: the
pair leaves the report entirely, so a difference swallowed inside one would be
spending that never appears anywhere. `Pair.Fee` is what left less what
arrived, and `report.Excluded` carries it out per currency alongside the legs —
named on the `money report` line and on its own row under the web report.
Nothing may pair on a mismatch without that difference reaching `Excluded`.
It is a fee and not a tag: no rule produces it, `report.ByTag` never sees it,
and it must not be turned into a synthetic transaction to make the total
reconcile. `Excluded` counts it only for pairs whose legs are *both* in view,
for the same reason it counts legs and not pairs — half a pair cannot say what
the other half received.

**`(transfer)` in a tag column is display only.** `Transaction.DisplayTag`
falls back to `model.TransferTag` for a paired leg so the column does not read
as a blank waiting for a rule, but nothing writes it: `rule_tag` stays what
rules.toml made it, `store.Tags` never returns it, and no rule can be built
from it. It is the same kind of label as `report.Untagged`. Do not "persist"
it — that is precisely the second verdict this design exists to avoid.

**A paired leg is not untagged.** `store.Filter{Untagged: true}` means "no tag
*and* no transfer", so a leg never turns up in `money ls --untagged`, the `u`
view or the rule builder's preview asking to be tagged — it is spoken for, and
the report drops it regardless. Covered by `TestPairedLegsAreNotUntagged`. An
*unpaired* leg is untagged like anything else, which is how it gets noticed.

**The pairing is derived state, exactly like the tags.** `Engine.Link` rewrites
the whole `transfers` table from rules.toml, so `money retag` re-derives both
halves of what that file decides. It runs over the whole index deliberately —
pairing within a filtered view would let a movement count as a transfer in one
report and not in another. Note the asymmetry with a schema *column*: a new
table is created by `CREATE TABLE IF NOT EXISTS` in `store.Open`, so adding one
does not force an index rebuild.

**Rules and transfers share rules.toml, so textual deletion is block-aware.**
`config.deleteBlocks` finds every `[[...]]` header, not only the kind being
deleted, because a block ends where the *next* block of any kind begins —
otherwise deleting a rule would swallow a transfer that follows it. Covered by
`TestDeleteLeavesTheOtherKindAlone`.

**Statements are the source of truth; the index is disposable.** Deleting
`index.db` at the root of the data directory and re-importing must reproduce
everything, with nothing lost.

**The index is a cache, so there are no migrations.** `store.Open` applies
`schema` and nothing else — no `ALTER TABLE`, no version column, no repair of
an index an older build wrote. Changing the schema costs one line in the
release note: delete `index.db` and import again. The two directions are not
symmetric, which is worth knowing before you assume something is broken:

- *Removing* a column leaves an older index still working, carrying the dead
  column and its data unread.
- *Adding* one the code reads breaks every command against an older index with
  `query transactions: SQL logic error: no such column: …` until it is deleted.

That asymmetry holds only because every statement names its columns —
**no `SELECT *`, and nothing may depend on column order.** Keep it that way.
Do not reintroduce migration machinery either; if re-parsing ever becomes too
expensive to ask for, that is a decision to revisit deliberately rather than a
helper to slip back in.

**Dedupe is by fingerprint**: `sha256(date | amount | normalised description |
ordinal)`, where the ordinal distinguishes identical lines *within one
statement*. Two identical purchases on one day both survive; the same line in an
overlapping statement does not duplicate. Unchanged files are skipped by
checksum before parsing at all — so an import with nothing new finishes in
milliseconds. That is the checksum skip working, not a failure.

**Money is `int64` minor units**, never a float. Per-account currency, no
conversion, and totals are never summed across currencies.

**The most specific rule wins, not the topmost.** `rules.Engine` sorts the
rules once in `New` and matches in that order: most literal characters first,
then fewest `*`, then account-scoped over unscoped, with a *stable* sort so
equally specific rules keep file order and the earlier one still wins. That is
what lets `*NIKOLA*` carve an exception out of `*NIK*` from anywhere in the
file, and a catch-all `*` sit wherever it reads best. A rule setting both
`match` and `type` requires both of them, and both count towards its literals.

The two orders must not be confused. `Engine.rules` stays in file order and
`Rules()`, `Usage` and `MatchIndex` all speak in file positions, because that
is what the rules screen numbers, what `config.DeleteRules` deletes by, and
what the user can point at in rules.toml; only `Engine.order` is sorted.
Anything new that reports a rule must report its file position too.

`config.AppendRule` still appends rather than prepends, but that now only
settles ties: a saved rule cannot displace an equally specific one written by
hand, while a narrower one is meant to take precedence and does.

**The report's period narrows the report and nothing else.** The screen opens
on last month, and `←`/`→` step along `report.Periods` — the named windows,
then the months the index holds. The report and the transaction list share a
scope (account, search, untagged), which `web.filterFrom` reads and
`state.filter` holds in the page; the period is read by the report handler
alone and kept in `state.report`, so opening on last month cannot hide the
rest of the index from the list. Nothing may put the period into that shared
scope, which is exactly what would make it leak. Covered by `TestReportPeriodLeavesTransactionsAlone`.

The axis is built from `store.Months` over the *whole* index, not from the rows
in view, or it would grow and shrink as the account or search filter changed and
move under the cursor. It is rebuilt on every request (an import can reach
further back), and the page names the period by its label rather than its
position, so a grown axis keeps the window the user was on. The windows are
relative to today, never to the newest statement: "last month" with nothing in
it reports nothing and says so, because silently answering for a month nobody
asked for is worse than an empty screen. That is also why an empty period and
an empty index give different messages — one asks for another period, the
other for an import.

**The report's sort rearranges rows; it never changes which rows there are.**
`report.Order` is passed to `ByTag` and is deliberately *not* part of
`store.Filter` or the shared scope — the period decides what is counted, the
order only how it is listed, and merging the two would make a sort able to hide
a tag. Currency stays the outer sort key under every order, because there are no
exchange rates to compare two currencies by, and every order falls back to the
tag so ties keep a fixed position instead of shuffling between reloads.
`Order.Column` is what lets the page mark the sorted heading without keeping its
own copy of that mapping; a new order needs a column of its own, which
`TestOrderColumnsAreDistinct` checks.

**The single-key shortcuts must not fire while typing.** The page's `keydown`
handler ignores them whenever the target is an input, or typing `i` in the tag
field would start an import. A view's own `key` hook runs first and is told
whether the user is typing, so a command *inside* a builder has to be a chord —
`ctrl+s` re-sorts the preview — because every printable key belongs to the
field being typed in.

**A rule's usage count is how many transactions it wins, not how many its glob
could match** — `Engine.Usage` counts by `MatchIndex`, so a rule shadowed by a
more specific one correctly reports zero. That is what makes the rules screen able to
find dead rules at all.

**A rule's `note` is documentation that round-trips.** It is a TOML key rather
than a `#` comment so `LoadRules` can return it, the builder can write it and
the rules screen can show it. It never takes part in matching — `rules.Engine`
does not look at it — and it must stay that way.

**`config.DeleteRules` and `config.ReplaceRule` edit rules.toml textually, never
by re-serialising the parsed rules**, because comments and formatting are not
recoverable from `[]Rule`. A rule owns the comment lines directly above it, so
deleting takes them with it; a comment separated by a blank line is a heading
for what follows and stays. Replacing keeps them — they say why the rule is
there, which editing its glob rarely changes — and rewrites only the rule's own
lines, leaving its *position* alone: position still breaks ties, so a rule that
moved could start beating an equally specific one it never used to. Every
writer goes through `writeFileAtomic`, and both editors re-parse the result
before replacing the file.

**An edit must not lose what the builder does not show.** The form has four
fields and a `Rule` has five, so `editRule` takes `Type` from the rule on disk —
never from the request — and the form says it is keeping it. Nothing may
round-trip a rule through those four fields alone — a pattern the user was
never shown is not a pattern they chose to remove. A new field on `Rule` needs
the same treatment or a field of its own. Covered by `TestEditKeepsTypePattern`.

**The rule builder's preview is what the rule is judged against, which is not
always "what is untagged".** For a new rule those are the same thing. For an
edit they are not: the rule's own transactions are tagged, so an untagged-only
preview would be empty for a rule that works. `previewGroups` adds them back,
and `rulePreview` keeps the ones the new glob stops catching on screen marked
`−` instead of dropping them silently, because giving one up is the decision
being made. Covered by `TestEditPreviewShowsWhatIsLetGo`.

**The web app is a frontend, not a second implementation.** Everything that
decides something stays in Go: amounts are formatted server-side (`money` is
text plus a sign, never a number the browser could add up), the rule builder's
preview is matched by `glob.Match` on the server rather than re-implemented in
JS, and the transfer preview runs `transfers.Analyze`. Keep it that way — a
JS glob or a JS sum is a second copy of a rule that would drift from the
first.
`static/app.js` only lays out what the API returns.

The server outlives edits to rules.toml in a way a one-shot command does not,
so `/api/retag` and `/api/import` re-read it first (import also re-reads the
account folders), and `/api/overview` reports `stale` when the file on disk no
longer equals what the engines hold. Every edit and delete sends the rule or
transfer it showed alongside the position, and is refused with 409 unless
rules.toml still holds exactly that there — a position from a stale page can
otherwise name a different rule. Covered by `TestStalePositionIsRefused`.

There is no auth by design (`--addr` defaults to loopback). Non-GET requests
must be `application/json`, which is what keeps a cross-site form from posting
to it; do not relax that without putting something else in its place.
That is why `/api/upload` takes files base64 inside JSON rather than as
multipart: multipart is precisely what a cross-site form can send.

**An upload only puts a file where `money import` looks.** It writes into an
existing account folder and then runs the same import as `/api/import`, so the
statements stay the source of truth and nothing reaches the index any other
way. It never overwrites a statement (identical contents are a no-op, different
ones a 409), refuses any name import would not read back — not a plain file
name, a dotfile, `account.toml`, outside the account's `include` — and checks a
whole batch before writing any of it. Covered by
`TestUploadNeverReplacesAStatement` and `TestUploadRefusesBadNames`.

## Adding a bank parser

Implement `parser.Parser` and call `parser.Register` from an `init`. Nothing
else changes; the name becomes usable in an `account.toml`. Use
`parser.ParseAmount` rather than hand-rolling decimal handling — it copes with
`1.234,56`, trailing minus, parenthesised negatives, currency codes and both
the ASCII hyphen and U+2212.

Keep PDF text extraction separate from parsing: the bank parsers expose a pure
`parseXText(text string, digits int)` so they can be tested against captured
`pdftotext` output without a PDF fixture. `pdftotext -layout` (poppler-utils) is
a runtime dependency of `nlb` and `traderepublic`; no Go library reconstructs
column layout as well.

It can be carried inside the binary instead: `scripts/build-bundled.sh` builds
a static musl `pdftotext` in a container and embeds it under `-tags bundled`,
and `pdftotextCommand` writes it to the user cache dir (named by content hash,
verified before reuse) and runs that. It is still the real executable run as a
subprocess, deliberately — not poppler linked in through cgo, which would cost
the pure-Go build, and not a different extraction API whose spacing might not
match what the parsers were tuned on. The tag is opt-in so `go build`/`go test`
never need the 5 MB file, which is gitignored rather than committed. Bumping
`POPPLER_VERSION` means updating its pinned SHA-256 too, and checking real
statements still parse — that output is what the parsers depend on.

Do not hardcode absolute column positions from a sample PDF. `pdftotext`
compresses runs of spaces, so columns shift with font and page size — derive
positions from the header line (`traderepublic`) or from the line being parsed
(`nlb`).

**The other side's account number belongs in the description, not in a field
of its own.** There used to be a `Counterparty` field; only `nlb` could fill it
honestly, `revolut` and `traderepublic` invented it with an IBAN-shaped regex
over the description, and the two disagreed on spacing, so one rule pattern
could not serve both. Now `nlb` appends its IBAN column to the end of the
description — at the end, and not in the position it held on the page, so an
IBAN wrapped across continuation lines stays contiguous for a glob to match.
A new parser must do the same rather than reintroduce a structured field.

## Verifying

```
go build ./... && go test ./... && go vet ./... && gofmt -l .
```

For end-to-end checks, build a throwaway data root under the scratchpad rather
than touching real data:

```
money --root /tmp/.../demo import
money --root /tmp/.../demo import          # must report 0 new
money --root /tmp/.../demo ls --wide
```

If you changed the schema, delete that root's `index.db` first. Nothing
migrates it, so a demo root left over from an earlier build either carries dead
columns or fails with `no such column`, depending on which way the schema
moved.

### Testing the web app

Drive the API through `Server.Handler()` with `httptest` — `newTestServer` in
`internal/web/server_test.go` builds a data root and an index tagged and paired
as an import would leave them, with "today" fixed so the report's default
window is predictable. That is how every web test works, and since the page
only lays out what the API returns, it is where behaviour gets tested.

For a look at the real page, run `money --root <demo> serve --addr
127.0.0.1:<port>` against a throwaway root. Delete the index first if you want
to watch an import do work, or it will be skipped by checksum and finish
instantly.

## Conventions

Comments explain why, not what. Errors name the file and the offending row or
key so a bad statement is actionable. Per-file import failures are reported and
the run continues; only a broken data root aborts.