"… und 5 weitere Ereignisse in diesem Zeitraum" was a sentence to read.
It is now a button: click it and the rest unfolds in place, click again
and the card goes back to eight rows.
Until now the only way to see the remaining entries was to narrow the
period or the kinds — which changes the question rather than the answer.
This happens in the browser, not through the URL, unlike the period and
kind filters. Those go through the address because a longer period brings
rows into play that were never loaded; here every entry in the period is
already on the page and the eight was purely presentational. A round trip
would mean waiting for data that is already there, plus a history entry
for something nobody wants to go back to.
The list moved into its own component so the state has somewhere to live.
KIND_LABEL and the item type moved with it, since they only describe this
list.
The button only appears when there is something to unfold, and once open
it offers the way back — otherwise the card stays long for the rest of the
session because somebody looked once.
Verified: 10 tests, five mutation-checked (ignoring the state, dropping
the way back, showing the button with nothing to unfold, losing the
singular, dropping aria-expanded each turn them red). 487 tests,
typecheck and build clean. Not seen in a browser — login goes through the
company account and the database is unreachable.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The bell is a shared pile: every active HR person sees every open note,
regardless of who wrote it. That was agreed and it stays the default —
this narrows it, it never widens it. You can now untick colleagues whose
notes you do not want to see.
What gets stored is the *exceptions*, not the selection. The difference
shows the day someone new joins HR: had the selection been stored, she
would be invisible to everyone until each person ticked her, and nobody
would notice her follow-ups piling up. This way she is visible from day
one and hiding her is a deliberate act. Same reasoning that made notes a
shared inbox in the first place — the silent gap is worse than a row too
many.
Own notes always come through: `note_mutes` rejects a self-reference, and
the predicate says so again rather than depending on a check constraint
staying put. Notes with no author come through too — hiding one because
nobody knows who wrote it is exactly the loss this list exists to prevent.
The rule lives in lib/notes.ts as one SQL expression because two places
need it: the bell in the header and the "Anstehend" card on the dashboard.
Two copies drift, and then the card counts something the bell does not
show.
No SQL function and no audit row, unlike anything that touches employee
data — this is a personal display preference, and an audit trail recording
every tick would make finding real changes harder. Same pattern as saved
reports and hire drafts, and the owner policy on note_mutes means a row
for someone else cannot be written even with invented values.
The checkbox flips immediately and flips back if saving fails; the list
gets clicked through several at a time and a round trip per tick feels
like hesitation.
Verified: 19 tests, five mutation-checked (or→and, dropping the own-notes
clause, inverting `not exists`, inverting the default, and losing the
email fallback each turn them red). Typecheck, lint, schema drift, 477
tests and the build are clean. Not seen in a browser: login goes through
the company account and the database is unreachable — the migration is
reviewed but has not been run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Beim Neuaufbau der Datenbank aus den Migrationen am 26.08. blieben
cost_centers, position_cost_centers, onboarding_tasks und
offboarding_tasks ohne Row Level Security: sie verlassen sich auf
den Ereignis-Trigger ensure_rls, den eine spaetere Migration erst
anlegt. Ein Ereignis-Trigger wirkt nur nach vorne.
Die Policies auf diesen Tabellen existieren und wurden nie
ausgewertet — PostgreSQL befragt sie nur bei eingeschaltetem RLS.
In jeder Aufstellung der Policies sah es richtig aus.
Der CI-Job prueft ab jetzt den Zustand nach dem Lauf, nicht nur
dass die Dateien durchlaufen. Genau in diesem Zwischenraum ist der
Fehler durchgekommen.
The dropdown is gone; each column header is now a link that sorts by that
column, with an arrow showing the direction. Clicking the column already
sorted reverses it; clicking a different one starts ascending again — going
from "Eintritt, newest first" to "Name" should give you names from A, not
inherit the previous direction.
Names sort by surname and then forename, as asked. Both parts reverse
together: turning only the surname would give Z-A across surnames but A-Z
within each one, which is visible immediately among the fifteen Aigner.
Three of the seven columns are not on the employee row. Bereich and Team
hang off the position, Standort off a lookup table, so they are fetched as
correlated subqueries rather than joins. That is not a style preference: the
same filter chain produces the page *and* the count, and a join onto
position_assignments would double every person who has held more than one
position over time — the line above the list would read 1,203 for 867 people.
Bereich is the level below the company, so it needs to walk up from the unit.
No recursion: org_unit_type has exactly four levels, so two hops up cover it.
Everything sorts `nulls last`, otherwise reversing the direction floats every
person without a position or location to the top.
The expressions live in lib/employee-sort.ts rather than in the page so the
generated SQL can be read in a test — the failure mode here is silent, the
list still shows fifteen rows, just the wrong ones. Eighteen tests, and the
rules are mutation-checked: dropping the forename, dropping the id tiebreaker,
dropping `nulls last`, sorting the location by its uuid, and shortening the
Bereich walk each turn them red.
Not seen in a browser: login goes through the company account and the database
is unreachable. Typecheck, lint, 458 tests and the build are clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Bei GitHub laeuft der Job auf dem Host und erreicht einen Dienst
ueber den durchgereichten Port. Der Gitea-Runner setzt den Job
selbst in einen Container: dort ist 127.0.0.1 der Container selbst,
und dort horcht kein Postgres — 'Connection refused'.
Beide Container haengen im selben Netz, also gilt derselbe Weg wie
in docker-compose.yml: der Dienst wird ueber seinen Namen erreicht.
Der Job 'Migrationen auf leerer Datenbank' brach beim ersten Schritt
ab: 'psql: command not found'. Das Abbild ubuntu-latest bringt keinen
Postgres-Client mit, die Vorbereitung der Datenbank ruft ihn aber —
deshalb lief der Job nie bis zu dem, wofuer es ihn gibt.
Genau diese Pruefung haette die fehlenden Rechte der Anwendungsrolle
gefunden, bevor sie beim Umzug auftraten: auf einer leeren Datenbank
scheiterte die Anwendung mit 'permission denied for table profiles'.
Beim Aufraeumen fielen zwei Regeln mit heraus, die Abzuege mit
Personendaten ausschlossen. Die Dateien liegen weiterhin auf dem
Server, ein 'git add .' haette sie mitgenommen — 857 Personalakten
mit SV-Nummern, Adressen und Angehoerigen.
Jetzt nach Ort statt nach Namen: /*.sql trifft die Abzuege im
Wurzelverzeichnis, laesst db/migrations aber unberuehrt (fuehrender
Schraegstrich). Damit greift die Regel auch fuer kuenftige Namen.
The schema/type drift check has been red in CI. It reported seven
discrepancies, and all seven were the checker's own fault.
The migrations put several clauses in one statement:
alter table employees
drop column if exists division_id,
drop column if exists team_id,
drop column if exists manager_id,
drop column if exists org_level,
drop column if exists is_lead;
The old pattern matched `alter table (\w+)\s+drop column (\w+)` as a single
regex, which finds exactly the first clause. So division_id was dropped from
the model and the other four stayed in it — the checker insisted four columns
existed that the OM cutover removed a year ago. The same cut the other way for
`add column`: kuendigungsschutz_bis, teilzeit_bis and aufenthaltstitel_bis are
each the second clause of their statement, so the checker never saw them and
called them typed-but-absent.
Now the statement is collected up to its terminating semicolon — with paren
depth tracked, so a semicolon inside a check constraint does not end it early
— and every clause inside is applied.
Verified by mutation, not by the green result alone: putting an invented
column into types.ts is caught, and reverting the parser to read only the
first clause brings back exactly those seven messages. That is the diagnosis
confirmed, not merely a passing run.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The database moved to a container of our own; the platform is gone.
This takes out what was left of it — and, where the leftovers were load
bearing, moves rather than deletes.
Moved, not deleted:
supabase/migrations/ -> db/migrations/ the schema's source of truth
supabase/build-org.ts -> scripts/build-org.ts
lib/supabase/types.ts -> lib/types.ts 52 import sites repointed
The bookkeeping needed care. It lived in `supabase_migrations.schema_migrations`,
and simply renaming the schema would have left the runner facing an empty
table: it would have called all 67 migrations pending and replayed them
against a database that is long since current. So the runner now creates
`migrationen.schema_migrations` and, once, copies the old rows across —
guarded so a second run does nothing and a fresh database skips it entirely.
Only then does migration 20260907100000 drop the old schema.
Deleted: the CLI config, the seed, the historical schema/function dumps
(nothing read them), scripts/umzug-von-supabase.sh (the move is done), and
both Supabase packages plus the CLI. Nothing in the application imported
them — the build now succeeds with no environment variables at all, which
is the proof.
Integration tests: six of them signed in through Supabase Auth and asserted
against the anon key and the service role. That model is gone, so the tests
were not portable — they are deleted. session-context and
employee-status-filter already ran on pg and are untouched; om-reporting is
ported to a direct connection because it guards a real risk (the reporting
line rule exists twice, once in SQL and once in TypeScript).
CI: the integration job started a Supabase stack. It now runs a postgres
service, applies deploy/db-init and every migration to an empty database —
that was the valuable part, and it still holds — then checks that a second
run is a no-op, which is what proves the bookkeeping works.
Docs: security-review.md audited a service-role key, a cookie adapter and
auth.users, none of which exist. Restating findings about removed components
would suggest today's system had been reviewed; it has not. It now records
what was removed and says a fresh review is due. data-model.md was already
marked obsolete and described the pre-OM schema; azure-migration.md was a
plan for a route not taken. Both deleted.
Verified: npm ci, typecheck, lint, 445 tests, build — all clean without the
packages. Integration tests skip cleanly with no database. Migration SQL and
the runner are reviewed but NOT executed: no Docker here, and the old
instance no longer resolves.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Sorting lives in the URL, not the browser. The list is built on the
server and fetched a page at a time, so a client-side sort would only
reorder the fifteen rows on screen — with 867 people that promises an
alphabetical list and delivers something else on page 2.
The sort key is the surname, because that is how the column reads:
"Aigner, Manuel", and whoever looks for someone looks under A. Postgres
runs with the Austrian collation, so Ö sorts with O rather than at the
end of the alphabet.
Only the surname reverses. The id stays ascending: it decides nothing
except ties, and it exists to keep the order total across page
boundaries. Reversing it too would still be deterministic but would flip
the fourteen Winklers relative to each other for no reason anyone asked
for.
The select sits in the filter bar rather than in a clickable column
header — a header would suggest it sorts what is on screen.
Verified: compiled SQL is `order by last_name desc, id` for Z-A; the
parse and direction rules are covered by tests that were mutation-checked
(breaking each rule turns them red). Not verified in the browser — the
login goes through the company account, and the Supabase instance no
longer resolves.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Der Umzug in einen eigenen Container hat zwei Luecken aufgedeckt:
1. Die Rechte von alpenwerk_app standen in keiner Migration — sie waren
auf Supabase von Hand im Dashboard vergeben worden. Auf einer leeren
Datenbank scheiterte die Anwendung deshalb mit 'permission denied for
table profiles', bevor die Anmeldeseite erschien.
2. docker-compose.traefik.yml lag nur auf dem Server. Ein frischer Clone
haette die Anwendung ohne Routing hochgefahren. Zusaetzlich haengt app
jetzt im Netz 'default', sonst findet es den db-Container nicht.
It was never used. The repository lives on a self-hosted Gitea, which
Vercel's git integration cannot connect to at all — so the documented
route amounted to "mirror to GitHub first", and nobody did.
vercel.json is gone, and with it the branch in next.config.ts that
switched off `output: "standalone"` when the VERCEL variable was
present. That branch was the only functional trace; everything else was
documentation and comments describing a second deployment path that did
not exist.
DEPLOYMENT.md loses its "two supported ways" framing and the whole
Vercel section — about fifty lines. Several statements next to it were
stale for a different reason and are corrected in the same pass: the
outbound-firewall table still listed Supabase's pooler (the database is
a container now, nothing leaves the server), the prerequisites still
demanded an existing Supabase project, and the .env table still asked
for a pooler connection string instead of the two new passwords.
The nightly job is described as what it is — a container in
docker-compose.yml — rather than as a replacement for Vercel Cron.
Migrations keep their references: two comments from July mention Vercel
Cron, and they describe what was true when they were written.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Supabase was only ever the host: the application has talked to PostgreSQL
directly through pg/Kysely for a while. So the move is mostly about
supplying what the platform used to supply.
Proved before building anything. All 65 migrations replay onto an empty
database, and the result matches production exactly — 183 columns, 25
policies, 68 indexes, 84 constraints, identical sets, no diff. The only
function missing from the rebuild turned out to matter, see below.
What the platform supplied, deploy/db-init now does:
- alpenwerk_app, explicitly NOBYPASSRLS. The whole access model is 21
RLS policies; a role that bypasses them would leave everything
working while showing too much, and nobody would notice.
- pgcrypto and pg_trgm. uuid-ossp was available on Supabase but is
used nowhere — no column default, no function calls uuid_generate_*.
- anon, authenticated and service_role as NOLOGIN placeholders. No
policy names them; they only carry grants the platform handed out,
and a data dump referencing them would fail to restore without them.
- A stub `auth` schema. The end state needs none of it — checked: no
foreign key, no policy, no column default refers to it. The June
2026 migrations do, and rewriting those would be falsifying history;
they describe what was true then.
The gap the comparison found: rls_auto_enable() and the ensure_rls event
trigger existed only in the running database, created by hand, in no
migration. That is the net which forces RLS on every newly created
table — the reason a forgotten policy yields an empty table instead of
an open one. A rebuild from migrations would silently not have had it:
everything works, and the next new table is unprotected. Now a migration
(20260819100000), verified by creating a table on the rebuild and
confirming RLS came on by itself.
Data moves separately, via scripts/umzug-von-supabase.sh: schema from
the migrations, then pg_dump --data-only --disable-triggers for the rows.
Without --disable-triggers every foreign key trips over load order. RLS
does not interfere — none of the 19 tables uses FORCE ROW LEVEL
SECURITY, so the owner writes through. The dump is deliberately left on
disk afterwards.
psql and node come from two `tools`-profile services rather than being
installed on the host, so the server needs nothing but Docker. The db
service publishes no port at all — reachable only inside the compose
network.
SUPABASE_DB_URL is renamed MIGRATE_DATABASE_URL, since after this it
describes something else entirely; the old name still works so existing
.env files keep running. Both were exercised, as was the error when
neither is set.
The deploy workflow is set to manual-only. Its preconditions were never
met — no secrets, and whether the job container can reach the host's
Docker daemon is untested — and failing on every push teaches people to
ignore red runs. It also needs updating for the new database service
before it could work at all.
Not verified: none of this has run in an actual container. There is no
Docker daemon on this machine. What is verified is the part that
decides whether it can work — the schema, on a real empty database.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The workflow asked for a `self-hosted` label. No runner on the instance
offers one, so the run would have sat in "Waiting" forever — no error, no
message, nothing to notice. The existing global runner elycon-runner-01
offers `docker` and `ubuntu-latest`, so both workflows now ask for
`ubuntu-latest`, the same label ci.yml already used.
That correction exposed a second thing the first version glossed over.
act_runner starts a container per job; mounting the Docker socket into
the *runner* does not put it in the *job*. Whether this job can reach the
host's daemon depends on the runner's config.yaml, which is not visible
from here — and the runner is global, so changing it affects every
repository on the instance, not just this one.
Rather than guess, the workflow now measures it in its first step and
fails with the fix if it cannot: which config lines to add for the
socket, or that SSH is the other way. Without that, the run would have
died three steps later on a message nobody could act on.
Both branches of the check were exercised: docker absent prints the
first message, docker present with no reachable daemon the second.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A push to master now builds and restarts the application on the server:
a Gitea Actions workflow on a self-hosted runner writes .env from the
repository secrets, applies pending migrations, rebuilds the compose
stack against the host's Docker daemon, and waits for the container's
healthcheck before calling the run green. Without that last step a
deploy counts as successful the moment the container *starts*, even if
the app inside it dies immediately.
Switching migrations on automatically turned up something that had to be
fixed first: supabase_migrations.schema_migrations did not exist at all.
Every one of the 65 migrations was unrecorded, because they have been
applied by hand all along. An automatic `db push` would therefore have
replayed all 65 against the live database — initial_schema and the OM
cutover included. The database was checked against a spread of
migrations first (it is at head), then baselined: all 65 recorded as
applied without executing them.
The runner is scripts/migrate.mjs rather than the Supabase CLI. It needs
only `pg`, which the project already ships, instead of downloading a CLI
whose version drifts independently of this repository; and it does one
thing — the missing files, in order, each in its own transaction — where
`db push` also diffs schemas and may do more than that. Bookkeeping goes
in the same table in the same shape the CLI uses, so `supabase db push`
from a workstation still works and still skips what already ran.
The workflow lives in .github/workflows, not .gitea/. Gitea reads
.gitea/workflows and falls back to .github/workflows only when the
former is absent — creating .gitea/ would have silently switched off
ci.yml, with the run simply never appearing.
Verified: both workflow files parse; the secret check names what is
missing and refuses; values starting with "-" or containing "=" survive
being written to .env; and the runner was exercised against the real
database with a throwaway migration — applied once, skipped on a second
run, and on a deliberate syntax error rolled back whole, recording
nothing. Both probes were removed; the count is back to 65.
Not verified: nothing has run on an actual Gitea runner — none is
registered yet. DEPLOYMENT.md §5 covers registering one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The checklist gets a print view: one A4 page, header carrying personnel
number, name, date of extract, entry date and position, the items in two
flowing columns so twenty-five fit, and two signature lines at the foot.
It prints the *state*, dated — not a blank form.
The default filename in the save dialog comes from the document title,
which is set to the agreed convention:
20260818_2884_Aigner-Manuel_Onboarding-Checklist
Surname first, like everywhere else in the application, so a folder of
these sorts by person and within a person by date. Umlauts are resolved
rather than stripped: the naive route (NFKD, then every non-ASCII to a
dash) turns "Müller" into "Mu-ller", because decomposition splits the
umlaut and the diaeresis becomes the dash. "Weiß" needs its own rule —
it has no decomposition and would otherwise vanish.
The reported defect: the printout carried the application's own top bar
— hamburger, bell, "Neueinstellung", sign-out. Those are controls; on
paper they are decoration, and on a checklist filed in a personnel
record, misleading. The rule now sits on AppShell rather than on this
one page, so the org-chart print view — which had the same problem —
gets it too, along with anything printed later. The shell's padding goes
with it: the type area is set by @page on the print page itself, and the
shell's would have been added on top.
The browser's own header line (date, title, URL) is separate — that is a
checkbox in the print dialog, not something CSS can reach.
440 tests pass, including 13 new ones pinning the filename convention.
Not yet seen in a browser.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The tab was keyed off employee.status === "Ausgetreten", and that stays
wrong for weeks: terminate_employee writes exit_date immediately no
matter how far out the date is, but only flips status once the date
itself arrives. A termination entered today for four weeks out left the
tab invisible for the entire notice period — exactly the stretch in
which IT access, hardware and deregistration actually get worked
through, and exactly where the checklist was supposed to live "next to
Onboarding," per the report that caught this.
The rule now reads exit_date instead: not null, and not a No Show
(which sets exit_date too, to the entry day, but never worked a day and
gets no checklist). rehire_employee resets both exit_date and
exit_reason to null, so a rehired person's tab still disappears the
same way it did before — nothing about that case changed, only the
signal the check reads.
Verified against the real database with the exact shape from the
report: a termination dated 30 days out. Status stays "Aktiv", the
checklist exists immediately, and the tab's own predicate says yes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The offboarding list was a checkbox fieldset inside the exit panel — four
items, never sent anywhere. Nothing in terminateEmployee's payload
carried them; ticking a box there recorded exactly nothing.
It's replaced with the same kind of list the entries got: its own tab,
appearing the moment an exit is recorded, with one item per row, a
comment on each, and — unlike the fieldset — a record of who touched it
and when.
The eleven items come from the same printed sheet as the entry list.
Nine are plain checkboxes. Two are text fields under "Vermerke":
remaining vacation and the balance transferred for payout — the sheet
names "Überleitung Salden für Auszahlung" twice, once as a task to do
and once as the actual figure, and those are genuinely two different
questions, kept as two items. Where the sheet still says "GKK" rather
than today's "ÖGK", it's left as written — that's the name the process
runs under internally, not a typo.
No Show gets no list. Never having worked a single day, there's no IT
access to revoke, no GKK registration to undo, no Dienstzettel to
collect — an empty checklist there would be a label with nothing behind
it. Both the tab and the auto-creation on exit check for this
specifically, not just the "Ausgetreten" status that No Show shares with
a real exit. Rehiring the same person hides the tab again — the data
stays, since it happened, but a checklist for someone currently working
has nothing to point at.
The engine (what counts as done, how progress is computed) moved into
lib/checklist.ts so onboarding and offboarding can't drift into two
different ideas of "done" the way two independent copies eventually do;
lib/onboarding.ts and lib/offboarding.ts bind it to their own item list,
and the tab UI is a single ChecklistPanel bound the same way.
Checked against the real database: a real exit creates all eleven items
in the same transaction as the exit itself; a No Show creates none;
rehiring flips the tab off while the old answers stay queryable. One
false alarm during that check turned out to be the user's own clicks on
a real employee's onboarding list, made in the browser while trying the
earlier feature — left untouched, not test debris.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The list existed on paper: one printed sheet per entry, twenty-five
boxes. What is on it is known only to whoever holds the sheet — it
cannot be searched, cannot be covered for while someone is away, and
says nothing about who ticked what.
Not every box on the sheet is a checkbox, and the differences carry
meaning, so the field kind is derived from the thing rather than
flattened:
Haken — the normal case. The Meldezettel is there or it is not.
Ja/Nein — Prämienanspruch had *two* boxes on the sheet, and that is
not decoration: "nein" is a finding, "not asked yet" is not.
One checkbox cannot say both.
Text — shoe, shirt and trouser size. The value is the point;
ticked off it would be worthless.
Every item takes a comment, and every item records who last touched it
and when — the part the sheet could never do.
Saved on click, not on submit. A checklist is worked through over days,
between other things; a save button at the end is where half a morning
goes missing.
The items live in lib/onboarding.ts, not in a table: a checklist is a
company process, not a master record. Stored per person is only the
answer, under the item's key — so an item dropped later leaves its old
answers standing instead of taking them along, and a file from back then
stays readable.
A list is created by hire and rehire, in the same transaction as the
hire itself: a hire without a checklist would be a half-recorded hire.
Rehire only adds what is missing and never clears an old tick — what
genuinely has to be redone is HR's call, and a program deciding it would
be guessing. People hired before this feature have no list and get a
button to start one.
Checked against the real database end to end: hire creates 25 open
items; checkbox, ja/nein, size and comment all land; a comment-only edit
leaves the tick alone; rehire tops the list up and keeps what was done.
The probe employee was removed afterwards — audit rows first, since the
log has no delete policy.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
There was none anywhere in the model, so no personnel-cost figure could
be produced at all, and an open position could not say whose budget it
would charge — which is the first question asked about a vacancy.
It hangs on the position, not on the person: the seat costs money even
when nobody sits on it. That is exactly the vacancy case. And not on the
org unit either, although it usually follows from one — a single seat
can be charged elsewhere (project, shared function) without the unit
moving.
As its own dated assignment table rather than a column, because
reassigning is an event with a date. Last year's costs have to stay
where they were incurred; as a column, every change would silently
rewrite every past report. Half-open [valid_from, valid_to), like
position_assignments and om_positions — in SAP OM this is A011.
25 cost centres seeded from the org tree: one per company, division and
department, with teams charging to their department, because a team is a
span of control and not a budget. All 823 positions were assigned from
their own start date, none left over. The number is the first five digits
of the org number, so it can be traced rather than looked up.
Reassignment refuses three things, each checked: the same cost centre
again, a switch on the day the current one started (that period would
never have been in force, and the range constraint says so), and a date
before the position exists.
Verified against the real data, which turned up a defect worth keeping:
a position that starts in the future is charged only from its start, so
asked about today it had no cost centre — and future positions are
exactly what the vacancy list is for. It is now read at the position's
own start date.
Two audit entries from the probe could not be deleted through the
application (the log has no delete policy — correctly), so I removed
them with the admin connection.
Still open, and the reason this is only the first of the three fields I
proposed: location and planned FTE.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adding the follow-up kind turned one of these tests green-for-the-wrong-
reason and one red: both had the three kinds written out by hand, so
"all of them are selected" no longer meant what the name said. That is
the failure mode a hand-copied list has — it does not break loudly, it
drifts.
The list now comes from ANSTEHEND_ARTEN, and the two cases that depend
on completeness build their input from it.
I committed the previous change with this test red. That was wrong; it
should have blocked the commit.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A note with a follow-up date is a task, and the overview is where tasks
are looked for. Until now it lived only in the employee file, which is
the one place you go when you already know who you are looking for.
Follow-ups behave differently from everything else on that card, and the
difference is the point: an entry on Monday is over on Tuesday, an
unfinished task is not. So there is no lower bound on the date — what
was due and never ticked off stays, marked overdue in red, sorted to the
top because it is sorted by date. A task that drops out of the list by
itself is a forgotten task.
Only "Erledigt" removes it. A note without a follow-up date never
appears: it is a record, not a task.
Checked against the live database — an overdue one and an upcoming one
appear, one without a date and one beyond the chosen period do not, and
ticking the overdue one off removes exactly it. The probe notes were
deleted again.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The app got slower as pages grew, and the reason was not the queries. It
was their number.
A transaction is pinned to one connection, and a connection runs queries
one after another. Every Promise.all in a withUser block looked like
concurrency and was a queue. Measured against the real database: the
round trip is ~36 ms, ten trivial `select 1` over one connection take
343 ms, over ten connections 39 ms. Nothing here is slow — the whole
dashboard payload is under 200 kB, and every table is around a thousand
rows.
More connections is the wrong answer: the RLS session context is per
transaction, so parallel reads mean parallel transactions, and those
multiply the connections the database will grant. Fewer round trips
instead. Postgres will return each sub-select as its own JSON column of
one result.
Per page view, counting the transaction frame:
shell (paid by every page) 10 → 4
overview 14 → 5
employee file 14 → 7
employee list 8 → 6
The overview plus its shell went from 24 round trips to 9 — about 860 ms
of pure waiting down to about 320 ms.
The one trap is documented where it bites: inside json_agg, Postgres
formats values itself and the driver's parsers (lib/db/pool.ts) never
see them. Dates, numerics and uuids come out identical; timestamptz does
not — "+00:00" where the driver gives "…Z". Timestamps are compared as
strings in lib/history.ts to decide what happened later, and those two
forms sort against each other wrongly. Every timestamptz in a bundled
query therefore goes through zeitstempel(), which was checked
character-for-character against the driver.
Four loaders moved out of their pages into lib/ so the number of round
trips can be measured without building a React tree, and so the new path
could be held against the old one field by field: same rows, same order,
same strings, for the overview and for four employee files chosen to
differ (with history, a chief, a planned entry, one with dependents).
withUser now counts the queries in each transaction and says so in
development past a threshold. Without that, this grows back: each new
tile brings its own query, and nobody notices until everybody does.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Sixty days and all three kinds was a guess, and it was the only one on
offer. Payroll cares about next month; the person filling a vacancy
cares about entries and nothing else. The card now takes a period and a
set of kinds.
The choice lives in the address rather than in the browser, because it
has to: the page is built on the server, and ninety days pulls in rows
that were never loaded at sixty. Filtering client-side would silently
cap the answer at whatever the first query happened to fetch. It also
means a filtered overview can be sent to someone and opened again the
same way.
Deselecting every kind returns to all of them. An empty card is not an
answer to a question nobody asked, and the way back would otherwise be
one click further than the way in. The default period and the full set
are absent from the URL instead of written into it, so a shared link
carries only what was actually chosen.
Anything the address cannot be trusted to hold is rejected: an unknown
period falls back to sixty rather than reaching the query, which would
otherwise be an invitation to ask for ten years of rows through a link.
Eight rows still, with a count of what did not fit underneath — this is
an overview, and the employee list is where lists belong.
Not verified in a browser: the built-in preview has no company sign-in,
so the page redirects to the login before it renders. Types, lint and
386 tests pass, and the filter's behaviour is covered directly.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The old refusal read: "Diese Abwesenheit ist noch nicht wirksam. Sie muss
über den Vorgang selbst abgebrochen werden." There was no such way. The
row sat in the file, the scheduled change kept running toward its date,
and nothing could stop either one.
That is not hypothetical. One person went absent in July, came back in
August, and still has a second return booked for the first of September
— recorded while they were already working again. The guard added
yesterday stops a third from being written; it does not remove the one
that exists.
Absences are called off whole, not field by field. For a planned
contract change the scheduled payload gets the affected fields lifted
out of it and runs on with the rest; an absence has no fields in that
map, and half an absence is not a thing anyone means. So the whole
scheduled change is cancelled, and what it had already noted on the
person goes with it: the date they were to be away from, the date they
were to come back on. Left behind, the profile would show an absence
with no event behind it. If the absence is still running, the return
date planned when it began applies again.
The link between the row and the scheduled change had to exist first —
start_karenz and record_karenz_return now record it. Existing rows get
it backfilled, but only where one running change of that kind falls on
that person and that day. Where two would match, the row keeps refusing:
guessing which process to cancel is worse than refusing to.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three things, all from the same screenshot.
The entry date can now be corrected. The Eintritt entry gets an edit
button — date only, no delete, because it is the start of the timeline
and a person without one has no beginning. Unlike every other entry it
needs no recorded before-values: the old date is on the employee row, so
this works on rows written long before any of this existed, which is
exactly the case that matters.
What hangs off that date is checked: no other event may precede it, exit
and absence start may not fall before it, and the first position
assignment moves with it — left behind it would leave days of employment
with no post, or a post with nobody in it. Someone already working
cannot be given a future entry date either; without that check a person
who has been here for years could be turned into a planned entry, and
the status derivation would agree.
That last rule came out of the rehearsal finding a hole: my first probe
picked a person with no other history rows, so the "nothing may precede
it" check had nothing to compare against and a date in 2099 sailed
through.
Second, the screenshot showed two returns from one absence, and the data
confirmed it: one person with two Rückkehr entries and a third still
scheduled, recorded while they were long since active. record_karenz_
return never checked that there was an absence to return from. Now it
does, and it refuses a second scheduled return — which would have
silently overwritten the first on its effective date.
Third, the history is filterable: upcoming versus done, a date range,
and the event types that actually occur in that file. The count of
upcoming items shows without filtering, because "what is coming" is the
usual reason to open the tab at all.
Still not deletable: Versetzung, Beförderung, Austritt, Wiedereintritt,
Reorganisation. Undoing those means restoring position assignments, and
that deserves its own step rather than being tacked onto this one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Employees from outside the EU, the EEA and Switzerland need a residence
permit, and HR needs to know when it expires — about eighty people in
the current data, across Türkei, Serbien and Bosnien. Two columns, the
same shape as the dismissal protection: a flag and a date that only
means anything with it. The date is optional, because an open-ended
permit has none and a mandatory field would force an invented one.
The nationality coupling deliberately stays out of the database. Putting
it there would mean keeping the country list in two places — SQL and
lib/countries.ts, where the picker needs it anyway — so an EU accession
would become a migration instead of a line in a list. Worse, correcting
somebody's nationality would fail the constraint while the old permit
was still attached, which is exactly the moment someone is fixing a
mistake. The UI decides whether the fields appear, and clears them when
the nationality moves into the free-movement area.
So the list is the load-bearing part, and it is tested: 31 entries, all
of them values the picker can actually produce, no duplicates, no third
countries. A missing nationality reads as "no permit required" — an
unanswered question is a reason to record it, not to demand papers.
The permit shows on the Stammdaten tab only for the nationalities it
applies to. A line reading "Aufenthaltstitel: Nein" under an Austrian
citizenship would look like information rather than a question that does
not arise.
Filter by it and by when it expires — the question behind that being
"whose permit runs out next quarter" — plus columns in the export.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A long-term absence recorded by mistake could only be undone by booking
a second event on top of it — leaving two entries in the file, the first
of which never happened. Karenz and Rückkehr can now be deleted and
corrected like the other entries.
For that to restore anything, the two operations first had to start
recording what they overwrote. start_karenz and record_karenz_return now
keep before/after the way change_employee_data does: status, kind of
absence, start, planned return — and for a return also employment type,
hours and the part-time variant. Without that there is nothing to revert
to, only a sentence.
The ordering rule HR asked for is enforced in the database, not just in
the UI: an absence cannot be deleted while a later return exists. A
return standing on its own would be a return from nothing, and the
person's status would derive from an entry whose starting point had been
deleted. Delete the return first and the absence frees up.
Rehearsed end to end on real data: absence recorded, return recorded on
reduced hours; deleting the absence refused; deleting the return put the
person back on Karenz with the original hours and the part-time variant
cleared; deleting the absence then put them back to Aktiv with no trace.
Rows written before today carry no before/after and stay untouchable,
with the reason they already gave. Planned absences are refused too —
they have their own operation, and their fields have no place in a
pending payload, which is why app_feld_karte carries a null group for
them rather than a plausible-looking wrong one.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Last step moved the four part-time arrangements out of the absence list
and recorded the reason in the history text. That answered "what
happened" but not "who is in one right now", and the profile showed
nothing at all. So it becomes a real field: teilzeit_art, with an
optional end date.
The objection I raised then still holds — a state goes stale, because
nobody goes back to note when a Bildungsteilzeit ended. teilzeit_bis is
the answer to it: with an end date a report decides for itself what is
still running instead of trusting that someone maintained the row. Left
empty it means "open end", which is an honest thing to say.
It runs through the ordinary change machinery rather than beside it. It
sits in app_feld_karte, so it shows up in the history as a field with
before and after, and can be corrected there like any other. The
description suffix from last step is gone — writing the same thing twice
is how two versions start disagreeing.
Reporting: filter by variant, by "in one at all", and by when it ends;
group headcount by variant, where the absence of one reads "Keine"
rather than a dash, because in a report that is an answer and not a gap.
Plus columns in the export and the import.
One gap found while rehearsing, and only because the probe happened to
pick a return date in the future: a scheduled return carries its payload
through pending_org_changes, and that payload did not include the
variant. Someone would have come back on reduced hours in April with the
reason gone. The daily run now carries it too.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Bildungsteilzeit, Elternteilzeit, Pflegeteilzeit and
Wiedereingliederungsteilzeit were offered as kinds of long-term absence.
Recorded that way, the person counted as absent: they dropped out of
headcount, their reporting line fell to a stand-in, and reports stopped
counting them — while they were in the building every week, just for
fewer hours. A part-time arrangement is not an absence; it is a change
of hours.
They now sit where they belong. Wiedereingliederungs- and Elternteilzeit
appear when recording a return from absence, as the reason someone comes
back on reduced hours — both typically begin exactly when the absence
ends. Bildungs- and Pflegeteilzeit appear under "Daten ändern" beside
the hours, next to the ordinary contractual change.
The reason is recorded with the change, not as a state on the person. A
state would have to be maintained, and nobody goes back to note when a
Bildungsteilzeit ended; a field that quietly goes stale is worse than
none. In the history it stands next to the value it explains, and stays
readable for good.
The check constraint on absence_type is deliberately untouched. Three
people carry the old values right now — two Pflegeteilzeit, one
Wiedereingliederungsteilzeit. Forbidding them would make existing rows
illegal. They are gone from the list of choices; the history stays
readable. Those three are worth revisiting, but that is a data decision,
not a code one.
Rehearsed against real data: an hours change with a reason and one
without, a reduced return with a reason and an unchanged one — checked
by reading both new history rows rather than "the latest", since now()
stands still inside a transaction and made an earlier probe report a
false negative.
I also overwrote tests/unit/absence.test.ts instead of extending it. The
original cases are restored; the diff is 49 added lines and 3 changed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Works council members, expectant mothers, parents on leave, registered
disabled employees, apprentices — each has its own rules that come
before a dismissal. The tool does not judge whether one is lawful, but
it must not stay quiet about it either, and looking it up on the
contract tab is exactly the step that gets skipped under time pressure.
So: a checkbox, an optional end date, and a red warning at the top of
the termination panel naming the date — or saying plainly that no end
was recorded. It shows for a no-show too; the protection runs from the
start of the contract, not the first day worked.
The date is optional on purpose. A works council mandate has a known
end, a pregnancy does not, and a mandatory field would force an invented
number. A constraint says only what cannot be: an end date without the
flag, which would be a leftover nobody could interpret.
The field goes the whole way through — hire, data change, contract
sheet, export, report criteria (as a yes/no and as a date range), and
the import. A field that exists in one screen and not the next is how
people stop trusting the numbers.
Terminating is now offered for planned entries as well, labelled "Nicht
angetreten", with No Show preselected. Without it a person who never
turned up stayed a planned entry forever, since nothing else can end
one.
One finding worth recording: tsc has been reporting success on a broken
program. A generated file under .next got corrupted when a build ran
against a live dev server, and its syntax errors suppressed semantic
checking everywhere else — two genuine type errors in this change went
unreported until I typechecked with .next excluded. The file is removed
and the ordinary typecheck is meaningful again.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Someone hired who then does not start needs an exit reason of its own,
and until now the case could not be recorded at all. Terminating on the
entry date failed on chk_assignment_range: the assignment was closed
with valid_to = valid_from, and an empty interval is forbidden there.
Moving the exit to the next day would have claimed a day of employment
that never happened — headcount, tenure, every as-of report.
"No Show" is now an exit reason, and it behaves differently in three
ways.
The exit date is always the entry date, whatever the caller passed. That
is what makes "never active" true rather than asserted: a person counts
as employed when their exit date is *after* the reporting date, and here
it never is. The status derivation needed no change at all — it already
says Geplant before the entry date and Ausgetreten from it on.
The position assignment is deleted rather than closed. The post was
never filled, it goes back to being open, and nothing records a holder
who never held it.
The status column goes to Ausgetreten immediately, even for an entry
still in the future. Otherwise it would read Geplant forever — nothing
runs later to correct it.
A constraint holds the first of those regardless of the path in,
including the import: exit_reason is distinct from 'No Show' or
exit_date = entry_date. "is distinct from" rather than "<>" so an empty
reason does not evaluate to null and slip through — the same three-
valued trap that let an earlier check pass the case it was written to
stop.
The dialog locks the date field when No Show is picked and says why, so
nobody types a date that would then be silently overridden. The
offboarding checklist is hidden: nothing was ever handed out.
Rehearsed against real data — a planned entry with a 2099 date passed
in, which came back as the entry date; derived status across three
reporting dates never Aktiv; a direct write with a mismatched date
refused; and an ordinary termination unchanged.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Employee names now read "Winkler, Hannah" wherever a person appears in a
list, a table, a heading or a tree node. That is the order a personnel
list is kept in, it is the order people are looked up in, and it finally
matches the sorting — the employee list has always been ordered by
surname, which made an alphabetical page look unsorted.
The name was being assembled inline in about twenty places. A rename
that catches half of them is worse than none, so it now goes through
fmtName in lib/format.ts and every display site calls it.
Sentences keep the natural order: "Hannah Winkler wurde versetzt" reads
like German, "Winkler, Hannah wurde versetzt" reads like a form. So the
toasts are unchanged and only labels moved.
Two things the change would have quietly broken:
The org chart's own filter matched against "first last". It now matches
either order, with or without the comma, so typing what you see works
and so does typing what you remember.
The print model sorted by the last word of the composed name, which
happened to be the surname and is now the first name — every printed
unit would have come out sorted by first name. It sorts on the surname
field itself now, which is what it meant all along.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The contract sheet had one field, "Merkmale", holding whatever applied,
comma-separated — and a dash when nothing did. Two problems in one row.
A dash cannot distinguish "has no company car" from "nobody ever
answered the question", and the entry read "Dienstwagen" without saying
which kind, which is the thing worth knowing since electric vehicles are
tracked separately.
Betriebsrat, Dienstwagen, laterale Führung and C-Level are now four
lines like every other line on the sheet, each with Ja or Nein. The
company car shows its drivetrain instead: E-KFZ or Verbrenner.
That label existed in three places — the dropdown, the hire summary and
now here. It lives in lib/dienstwagen.ts, so the same car cannot end up
named differently depending on which screen you are looking at.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Deleting and correcting a history entry stopped at the present: anything
not yet effective stayed put. That was not a principle, it was a missing
link. A planned change lives as a payload in pending_org_changes, and
nothing tied it to the history row — only a person and a date, and the
data already holds an Eintritt and a Vertragsänderung sharing one. So
employee_history now carries pending_id, set by change_employee_data
when it schedules something.
One planned change can carry two history rows: Stammdaten and Vertrag
are kept apart but scheduled together. Taking one back therefore strips
only that group's fields from the payload, and cancels the operation
only when nothing is left. Correcting one rewrites its group and the
effective date, and touches no employee data — the change has not
happened yet.
An entry stays on its side of the present. Pulling a planned change into
today, or pushing an effective one into the future, would mean adjusting
the employee record and the pending payload in opposite directions;
that is what the real operations are for.
Existing rows were linked where exactly one running operation matched
the person and date and no other row had claimed it. All five of them
matched. Anything ambiguous would have kept the old refusal, which now
says the actual reason.
The edit dialog surfaced a bug in useDialogFocus that predates it: the
effect depended on the identity of onClose, which almost every caller
rebuilds on render, so it re-ran after each keystroke and its cleanup
pulled focus back to whatever opened the dialog. Any dialog with a text
field would have accepted one character. It never showed because until
now no dialog kept its own state next to its own onClose.
Rehearsed against real data: a two-row planned change corrected, one row
taken back with the operation continuing on the rest, the second taken
back with the operation cancelled, and both refusals.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
"Winkler M" returned four people, two of whom are not called M: Karin
Winkler is a Montagemitarbeiterin and Katharina Winkler a
Maschinenbedienerin. The job title was searched with the same weight as
the name, so a single letter matched the start of a job word just as
readily as the start of a first name.
Searching job titles is worth keeping — "dreher" finding the CNC-Dreher
is useful. So the search is now tiered: names alone first, and the job
title joins in only when the names return nothing at all. A minimum word
length would have been the simpler rule, but any threshold is a guess;
this one is decided by the data in front of it.
Checked against the live data: "winkler m" gives Martin and Magdalena,
"winkler h" Hannah, "dreher" and "montage" still find their trades, and
"winkler montage" finds Karin Winkler — no name matches both words, so
the fallback does what was meant.
When the fallback runs, the result line says so. Without that, a list of
people whose names look nothing like the query reads as though the
search invented them.
Costs one small count query, and only when text was typed.
Not verified with next build: a dev server from an earlier session is
holding .next, and the user is testing in it. tsc, eslint and 297 tests
are green, and the search itself was run against the database.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Searching "Winkler H" returned all seven Winklers instead of the one
Hannah. Each word was matched as a substring, so "H" hit T-h-omas,
Kat-h-arina and CNC-Dre-h-er:in — every row. The shorter the input, the
more useless the result, and an initial is the shortest input anyone
would type.
A word now has to match at the start of a word: either the haystack
begins with it, or a space does. The haystack is first name, last name
and job title joined, with hyphens, slashes, colons and dots flattened
to spaces, so "dreher" still finds CNC-Dreher:in and "cnc" still finds
both the Dreher and the Fräser.
Checked against the live data before and after: "winkler h" now returns
Hannah Winkler alone, "h winkler" the same in either order, "winkler
kat" the two Katharinas, "dreher" the twelve CNC-Dreher.
The trigram index on the concatenated name no longer applies, which is
the price. At under nine hundred rows the scan is a few milliseconds; an
index on the same expression brings it back when that stops being true.
LIKE's own wildcards are escaped now — typing "100%" searched for
everything before.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every deletion of a Vertragsänderung failed with "new row for relation
employees violates check constraint chk_weekly_hours". The revert wrote
one UPDATE per field, and chk_weekly_hours ties two of them together:
Vollzeit means exactly 38.5 hours, Teilzeit means something in between.
Setting the employment type back to Vollzeit while 37 hours still stood
produced precisely the state the constraint forbids. It hit nearly every
contract change, because the form changes those two together.
Collecting the assignments and writing them in a single UPDATE removes
the intermediate state entirely. The state being restored was valid once
— it is in the history because it was — so restoring it whole is safe.
A violation can still be real: if a later change touched one of a
coupled pair on its own, the old value no longer fits today's state.
That case is caught and reported as a sentence instead of surfacing a
database error in a toast.
My tests did not catch this, and could not have: the revert lives in SQL
and the suite has no way to run it. What did catch it was HR clicking
the button. The rehearsal script now covers the reported case, an
unrelated single-field revert, and the genuine conflict.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
HR can now delete a history entry, but only where deleting one is an
honest thing to do — and deleting it also undoes it.
The rule they asked for is the interesting part: the last valid change
wins. Deleting an entry walks its fields one at a time. If a later entry
touched the same field, the current value stays — that later change is
the one in force. Otherwise the field goes back to what the deleted
entry recorded as its "before". So the middle of three entries can be
removed without an old value overwriting a newer one.
Four kinds of entry refuse to be deleted, each saying why in the place
the button would have been. Eintritt anchors the timeline. Transfers,
promotions, absences and exits moved positions and status — they have
proper operations for that, and guessing backwards is how you corrupt an
org chart. Anything not yet effective hangs off a planned change, and
that link is not trustworthy: there is no key between a history row and
its pending row, only a person and a date, and the data already has an
Eintritt and a Vertragsänderung sharing one. Matching on the date would
eventually cancel a change nobody meant. And entries from before the
history carried values have nothing to fall back to.
Confirmation is not "are you sure" — that question gets a reflex yes by
the third time. The dialog says what will be different afterwards: which
field goes back to which value, and which one stays because something
later claimed it.
employee_history keeps its append-only policies; delete_history_entry is
SECURITY DEFINER and checks the permission itself in its first line. The
audit log keeps the deletion with the values that were removed, and the
audit log genuinely cannot be edited.
The rule lives twice — in SQL and in lib/history.ts. The database is the
authority; the copy exists so the UI can hide a button that would fail
and print the reason instead. Rehearsed against real data in a
rolled-back transaction first: the later change held, the untouched
field reverted, all four refusals fired.
Also corrected in the data catalogue: I had written that
require_hr_admin was called by nothing. It guards all sixteen mutating
functions.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
HR reported it from testing: change someone's address and their history
shows "Geänderte Felder: Adresse, Ort" — the new address is on the
Stammdaten tab, the old one is nowhere. It was recorded, but only in the
audit log, which is a different page sorted by time and actor rather
than by person. So you had to already know what you were looking for to
find out whether an address had ever changed, let alone what it used to
be.
The field-by-field diff was being built anyway and written to the audit
log. employee_history now carries the same list, and the person's
history renders it as an expandable Feld / Vorher / Nachher table — the
same table the audit log uses, lifted into a shared component so the two
views don't drift into reading differently. It expands with <details>,
so the values are in the page: findable with Ctrl+F, present when
printed, no script involved.
The duplication with audit_log is deliberate. A person's history should
be readable on its own, including after the log is eventually thinned by
a retention rule.
Rows written before today stay without values. They could only be
reconstructed from the audit log, and the link is not reliable — no key,
only a timestamp and a person. Honestly empty beats plausibly wrong.
The migration was generated from the live function definition rather
than retyped, and the diff is four lines: two column lists, two value
lists. It carries a self-check that raises if either insert failed to
pick up the new column, and it was rehearsed inside a rolled-back
transaction against real data first — the probe confirmed the old street
name lands in the history row.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Opening the org chart PDF preview locked up the browser tab. The
measurement that fits each sheet to the page fed itself: the effect
listed onFaktor in its dependencies, and onFaktor was an arrow function
created fresh on every render, so the effect re-ran after every render.
It measured, reported the scale, and the report called setState with a
newly built object every time — new object, so React saw a change,
re-rendered, and the effect ran again. Measure, render, measure, until
React gave up with "Maximum update depth exceeded".
Two changes, and the mutation test says either one closes the loop on
its own: the callback now lives in a ref so the effect depends only on
the sheet identity, and the reducer returns the previous state unchanged
when the scale has not moved. Both are worth keeping — the ref stops the
effect from re-running, the guard stops pointless renders.
This shipped broken, and the reason it shipped is in the test file now.
Every element in jsdom is zero pixels, so the measurement bailed out on
its first line and the feedback never started; nine tests covering the
selection, the page count and the hierarchy all passed against a
component that froze on contact with a real browser. The new test gives
the elements a size, and fails with the exact error a user hits.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Six comments and doc lines put the number of RLS policies at 58. It is
21 — counted from pg_policy while building the data catalogue. The
figure appears in load-bearing prose ("all 58 policies call
is_hr_user()", "all 58 policies stay unchanged"), where being wrong by a
factor of three invites someone to go looking for the missing thirty-
seven.
The two occurrences inside supabase/migrations/ stay as they are. That
file already ran against the database; its comments record what was
believed at the time, and editing them would make the file differ from
what was applied for no gain.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every employee address in the database sat on test.manner.at, a domain
that reads like a company one. employees.email is the *private* address,
so an address shaped like a company mailbox invites being taken for one
— and eventually being written to. All 856 rows now sit on
privat.alpenwerk-test.at, rebuilt from first and last name, and the seed
generates the same domain so a reseed does not bring the old one back.
Umlauts are spelled out the way they are here (Höller becomes hoeller),
other accents are flattened, and where two people share a name the
personnel number is appended.
The first attempt got this wrong in a way worth recording. It wrote
ma<number>@ for all 856 rows instead of the intended name form, and the
check I had built only asked whether the results were unique and
well-formed — which they were. Two defects, both invisible to that
check: '\.+' inside a SQL literal was read as "any character, one or
more" and collapsed the whole local part to a single dot, and the
replacement string for the accent mapping had one character too many, so
the mapping was shifted. The fix uses '[.]+', a character class needing
no escape at all, so it no longer depends on how the connection treats
backslashes.
Untouched on purpose: app_users.email and profiles.email are the sign-in
accounts, and rewriting those would lock people out.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The one document describing the schema, docs/data-model.md, predates two
rebuilds. It names divisions/departments/teams and a positions table that
no longer exist, describes Supabase auth with an anon key and a service
role that were removed, and puts the policy count at 58 when it is 21.
Anyone reading it to understand the data would have been misled on every
count.
docs/datenkatalog.md replaces it, and was not typed up from memory: the
columns, defaults, keys and check constraints were read out of
information_schema and pg_catalog on the running database. Fifteen
tables, 142 columns, ten enum types, 21 policies. Where a rule appears in
prose, the constraint it comes from is named next to it.
Some of it only became visible by asking the database rather than the
migrations. generate_company_email and the is_hr_admin pair are still
defined but nothing calls them any more. Position numbers look like a
six followed by seven digits because the generator builds them that way,
not because anything enforces it — the column requires only uniqueness.
monthly_salary_gross is dead weight kept in case old rows hold data.
Three claims I drafted were wrong and the database said so: the position
number format, the event trigger's name (ensure_rls, the function behind
it is rls_auto_enable), and which tables deviate from the plain
is_hr_user() policy.
The old document keeps a pointer at the top instead of being deleted —
it is linked from the security review, and a stale document that says so
is more useful than a dead link.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The export offered four criteria — unit, location, status, employment
type — while the employee record carries around twenty selectable
attributes. Anything else had to be filtered by hand in Excel afterwards,
which is how a payroll hand-off stops matching the application it came
from.
All of them are now filters: contract type, blue/white collar,
collective agreement, paygrade, internal/external, gender, company car
and its drivetrain, works council, lateral leadership, C-level, type of
long-term absence, weekday worked, dependents on file, and open ranges
for entry, exit, birth date and weekly hours. The unit filter covers
every level rather than only divisions, so a single department can be
selected without going the long way round.
They live in one table in lib/report-criteria.ts, which the filter panel
builds itself from, the parser validates against, and the query turns
into conditions. A new criterion is one entry there and nothing else —
and it cannot end up working in the report while being silently ignored
by the export.
The two export links and the saved-report config now carry the query
string through as it stands instead of listing the parameters they know
about. That enumeration was the actual defect: adding a filter meant
remembering three separate places, and forgetting one produced an export
that quietly disagreed with the figure on screen.
Validation is not housekeeping here. These values reach SQL comparisons
and the download filename, i.e. a Content-Disposition header; what is not
in the list does not get through.
The company car dropdown leaves the employee list. It is one of twenty
equals under Berichte now, where the selection can also be exported —
which was the point of asking in the first place.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
