The calendar mixed two kinds of truth in one table. `planned_leave_day` was
simultaneously the allocator's output cache, the user's override store
(`is_pinned`, plus ad-hoc rows with no `wish_id`), and the resource ledger that
analytics and the age-4 maths read. The planner therefore had to ask "am I
allowed to touch this row?" on every sync, and got it wrong at the edges.
Every user assertion is now one `day_assertion` row — a concrete `day_type`
replaces a pinned day, a row with no `wish_id` replaces an ad-hoc day, and
`day_type='none'` replaces `planned_day_exclusion`. `planned_leave_day` becomes
purely derived with a single writer, and `PlanningService.pinned_keys()` is the
single answer to "is this day pinned", used by the allocator's loader, by
`sync_plan` and by the read router alike.
On top of that, every calendar band is now a real event: an ad-hoc band opens
the same full editor as a wish-backed one and saving it promotes it to a
`LeaveWish`, so it can be named, repeated yearly, assigned to either or both
parents, and handed to the planner. "Planering" is a `moveable` toggle plus a
from/to month bound plus a separate lock/unlock action on the days themselves,
replacing a three-rung ladder whose "planner decides" rung set a search window
equal to the drawn dates and so could never move anything.
Bugs found and fixed along the way:
- The plan erased its own history. Fixed windows clip to today, so an elapsed
auto-placed day fell into `stale_keys` and was deleted on the next sync —
and a background sync runs on every workspace load. Past rows are now frozen
and counted as spent.
- Migration 010's downgrade silently destroyed every pinned day and exclusion
(`is_pinned` returned as DEFAULT 0, the exclusion table came back empty,
exit 0, no warning). It now reconstructs both from `day_assertion`, and 009
itemises what it cannot recover before dropping the table.
- Deleting a period left its assertions behind, locking those dates forever
with no UI path to clear them. Wish deletion now clears them by `wish_id`.
- `DayAssertionService` and `LeaveWishService` committed multi-statement writes
without a transaction; a crash mid-write diverged the tables or half-deleted
a series. Both now run in a unit of work.
- A sibling year's shortfall was never surfaced: the toast only inspected the
edited occurrence's outcome, so a fully unfunded later year rendered as an
empty band with no warning.
- A flexible window whose bound had entirely elapsed was dropped from the
proposal with no outcome at all. It now reports as `unplaced`.
- Unlocking a period deleted its days instead of returning them to the planner.
- The e2e global setup posted to a route deleted in this change, which would
have failed the whole suite before its first test.
Gates: backend 938 tests / 91% coverage, frontend 1351 / 97.4 | 92.2 | 94.37 |
97.4. Migrations 009 and 010 audited against a hand-built realistic database:
the upgrade reconciles exactly and a constraint violation fails safe, leaving
the schema at 008.
Not complete. Step 4's toggle, month picker and lock/unlock action are verified
by browser walk and unit tests, but the partial-lock readout was not exercised.
`ideas.md` §7 adds a full plan for the vacation-forfeiture defect (the planner
detects expiring days and then does nothing about them) — not implemented here.
Remaining smaller items are listed in the session's restructure notes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T4wjgbVL4HAC5Rdaw75VbV
A yearly-repeating period kept the same calendar date instead of the same
weekday: 2026-10-05..10-11 (Mon-Sun) recurred as 2027-10-05..10-11 (Tue-Mon)
rather than 2027-10-04..10-10. The snap added in 224fe7a only ever covered
window_mode == 'fixed'; 'flexible' wishes were unaligned at both stages.
project_window now snaps both modes. A flexible bound is still a bounding
range, but it keeps its weekday alignment year over year with its span held
fixed, so the search space never shrinks and a bound that is exactly as long
as the target (its only candidate) repeats week-shaped.
_rank_candidate gains a Monday-start term between cost and start, so equal-cost
candidates break on week alignment instead of "earliest". It sits after cost
deliberately: a cheaper non-Monday start still wins, because red days outrank
alignment.
_reuse_already_funded_window would otherwise freeze a pre-existing misaligned
placement forever, so it now releases a year whose reuse winner is misaligned
when a fresh search ranks strictly better -- never when one of its days is
pinned or excluded, so a deliberate per-day override always survives.
Migration 008 realigns flexible series already materialised under the old rule;
projection runs once, at create/update, so nothing repairs them on read.
Alongside this, the planner now says when it moved a period to catch red days.
WindowOutcome carries shift_gain_days (the saving over the naive bound-start
placement) and the calendar toasts it, once, on the sync that decided it.
A fixed window is never auto-moved -- its "exact requested placement,
unconditionally" guarantee stands -- but its shift suggestion is no longer
hidden behind a shortfall, so a fully-funded one can surface a cheaper week too.
Editing a period in the popover now also sends window_mode: 'fixed' with a
matching target. Preset chips create flexible wishes; dragging one down to an
exact week left it flexible with a week-long bound, which is the degenerate
single-candidate case that produced the reported dates.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DXPBfr1v7F71t6QcbHCaaT
Three strands of work that grew together across the same settings screen and
API client, so they land as one change.
Per-year preschool planning days ("studiedagar"): new table, repository
(SQLite + Postgres), service, router and migration 007, plus the hooks and
Settings UI to plan and record them per child and year. Allocation now
reserves the recorded figure for a year where one exists and only falls back
to the child-level estimate otherwise, marking that fallback as assumed.
Partner invitations: workspaces issue and list invitations, registration
accepts an invite token to join the inviter's workspace instead of creating a
personal one, and a /join/:token page plus the settings invite panel drive it.
Vacation forfeiture: tips and the dashboard now distinguish days a parent
genuinely loses from days that merely age, so the warning stops firing on
balances that were never at risk.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The scraper never read "överlåtbara dagar" from Försäkringskassan. It only
visited ansok-och-planera-foraldrapenning#/valjdagar and derived the
reserved/transferable split in build_snapshot, assuming transferable days
are spent before reserved ones. On real data that gave 49 for the user and
0 for the partner where FK states 129 and 71.
The assumption cannot be repaired: reserved-first, transferable-first,
proportional and cap-based models were all checked against real figures and
none reproduces FK's accounting (reserved-first fits the partner but is 10
days off for the user). FK publishes the real numbers per parent per tier on
the min-foraldrapenning page, which the scraper had never visited.
- Scrape the "Dagar ni kan föra över" panel after the valjdagar extraction,
summing sjukpenningnivå + lägstanivå per parent. The totals are not
literals on the page. Parse only direct-child <p> elements: the first
group embeds a tooltip with decoy "45 dagar"/"90 dagar" text.
- build_snapshot accepts the reported value and derives reserved days from
it; the old arithmetic remains only as a fallback. An implausible FK value
degrades to that fallback rather than failing the fetch.
- Record provenance in transferable_days_reported (migration 006) and badge
derived values as estimates in the dashboard breakdown. Drop the footnote's
claim that at most 150 days are transferable — real data contradicts it as
a per-family figure.
- Select the child's section by name when FK lists several; never guess.
Verified against a live BankID scrape: scraped 240/101/139 with 129
transferable, deduced partner 240/169/71 with 71, both flagged as reported.
Incidental fixes found along the way:
- The fetch router never forwarded the partner's transferable figure to
persist_partner_deduction, so the partner's stored snapshot would have kept
the flawed split even once the parse landed.
- 'deduced' snapshots were labelled "Manually entered" by a two-way ternary
over a three-value union. Replaced with a compiler-checked exhaustive map.
- Alembic resolved its database from alembic.ini alone, ignoring
APP_SQLITE_PATH, so migrations silently targeted the developer's real
database even when pointed elsewhere. env.py now resolves it, preferring
-x sqlalchemy.url, then a real APP_SQLITE_PATH, then the ini value.
- backend/captures/ was not ignored: the root pattern was anchored as
/captures/ despite a comment saying scraper dumps must never be tracked.
- CLAUDE.md quoted test counts and coverage percentages that had gone stale,
which made a correct verification report look suspect. Gate thresholds kept.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DXPBfr1v7F71t6QcbHCaaT
Taking one day out of a period the planner placed did nothing that
lasted: the period still asked for a full week, so the next sync put the
day straight back. The option was disabled rather than lying about it.
A removed day is now recorded as its own thing — no leave for this
parent, this child, this date — and the planner honours it on every
later sync. It is kept apart from the working-day overrides, which
answer a different question about the calendar rather than about the
plan.
The day belongs to the parent whose band was clicked, so removing one
parent's day leaves their partner's alone; where both parents were to be
off on one funding arrangement, the partner is simply funded on their
own instead. A removed day shortens what the period is asking for rather
than counting as a day the plan failed to find, so the advice about
closing a shortfall does not start arguing for a day nobody wants.
Nothing is removed without a way back: the toast undoes it, and the day
itself is marked in the calendar and offers to restore.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ei7gCZzdd751uZtPzb1Pep
A repeating period was a single row that the planner expanded into one
window per year at read time, so no year had anywhere to differ from the
others. Editing one year rewrote them all, and removing a year in the
middle had to split the wish in two. Neither survives contact with the
thing this app is for: Easter moves, and the cheapest week is not the
same week every year.
Each year is now its own row, linked by a series id. Editing one year
touches one row; editing the series re-projects the rest from it.
Removing a year is one delete. The weekday snap is worked out once, when
the rows are made, and stored, so a row now says what it means rather
than being re-derived on every read. Recurrence stops being a stored
field and becomes what it always was in practice — an instruction at
create and update time; whether a period repeats is whether it belongs
to a series.
Migration 004 rebuilds the table and expands the existing rows, keeping
each original row as its first year so the days already planned against
it still point somewhere. A Feb 29 anchor recurring into a common year
used to raise out of the allocator; it clamps now.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ei7gCZzdd751uZtPzb1Pep
Asking for a full week off never spent anything on the Saturday, the
Sunday or a holiday, and a test now says so outright rather than leaving
it to be rediscovered. Those days are free because vacation cannot be
taken on a day that is not a working day.
Parental benefit is a payment rather than a schedule, though, so it can
be drawn on a red day to turn an unpaid day off into a paid one — at the
cost of a day the family would otherwise keep. That is now a workspace
setting, off by default, and it never reaches for a vacation day. With it
off the planner behaves exactly as before.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Ei7gCZzdd751uZtPzb1Pep
Manual entry now takes one parent's sjukpenningnivå, lägstanivå and
dubbeldagar left (plus optional reserved days) and derives the rest in
build_snapshot, shared with the scraper. The scraper also reads the
family panel and exposes a partner deduction; with two parents it is
saved as a 'deduced' snapshot (migration 002). Parent count is the single
source of entitlement: one member gets 480 days and 96 savable at age 4,
two get 240 and 48 each. Workspaces are capped at two parents.
Also lands the in-flight squash of migrations 002–011 into the initial
schema and the removal of planning_mode.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Security
- Every route now requires a login and refuses ids outside the caller's
workspace with 404: plan, leave-wishes, day-overrides and vacation-days
were fully anonymous; every other router trusted client-supplied
workspace/child/parent ids. Scraper sessions are bound to their owner.
- Invitations can no longer be created for a foreign workspace, and a
redeemer who is already a member (or a full workspace) is refused.
- Production refuses the default JWT secret; API docs are hidden in
production; a failed migration aborts startup instead of serving.
- Login and register no longer reveal whether an email exists; inactive
accounts are rejected; /auth/refresh is rate limited.
- The FK scraper writes debug artefacts only when explicitly enabled,
logs no scraped values or page text, clears the BankID QR as soon as it
is scanned and purges finished sessions after five minutes.
- Real FK page captures are removed from the tree (history rewrite
follows); the parser tests use a synthetic fixture.
Performance and correctness
- Plan apply/sync run inside one locked transaction with batched writes
instead of one commit per row, so a failure mid-batch leaves nothing
half-applied and concurrent syncs no longer corrupt each other.
- SQLite runs in WAL mode with a busy timeout; 14 indexes cover the
filters that were full scans; argon2 hashing runs off the event loop;
responses are gzipped.
- Frontend: a 401 now refreshes the token once and retries, so sessions
no longer die silently after 15 minutes; every data hook discards
stale responses; logout clears state even when the server call fails.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01CsXCCcDfmV4Sps4ZQSvNEY
Three faults, all invisible while the tests passed. Advice was built for the
first child only, so a second child got none. The planned days feeding every
count were fetched once per parent and appended to one list, so a couple's
days were counted twice and every intensity and projection derived from them
was wrong. And dismissing a tip forgot the moment the page reloaded.
Tips now carry the date they concern, lifted from figures the rules already
computed, so a view of a month can show only what applies to it. They are
built per child and identified per child, so the same rule firing for two
children is two separate pieces of advice.
A dismissal is stored against a fingerprint of the facts that were dismissed.
When those facts change the advice returns with its new numbers, rather than
one click silencing a warning about a forfeiture that keeps growing.
Three changes that together unblock treating a calendar period as one
editable object.
Migration 008 adds `planned_leave_day.wish_id` (nullable, ON DELETE SET NULL)
so a persisted day knows which `leave_wish` produced it, `is_pinned` to mark a
day the user hand-edited, and `leave_wish.target_days` so a period drawn on a
calendar need not be a whole number of weeks. Apply-time identity still keys on
day_type/is_dubbeldag only, so a wish_id-only difference is a metadata update
rather than a collision.
The age-4 savable ceiling was applying the per-child figure of 96 to each
parent's pool, modelling a two-parent family as able to save 192 days past the
child's fourth birthday. `resolve_age4_savable_cap` is now the single place
that decides 48 (couple with two parents) versus 96 (everyone else), and both
AnalyticsService and the allocator read it.
Migration 009 drops `workspace.planning_goal` and `leave_wish.goal_override`.
The allocator's tier-3 policy collapses to one rule; `salary_type` still
branches, since that reflects what a vacation day actually costs a parent
rather than a user preference.
Preset wishes were created with `name: t(preset.labelKey)`, so the
translated string was persisted as the wish's name. A wish created while
the UI was Swedish stayed "Jul" in English forever -- the Swedish text
was the data, so no display-layer change could fix it.
Wish names are user-editable free text, so `name` cannot simply be
translated at render time. Migration 007 adds a nullable
`leave_wish.name_message_key` recording where a name came from: preset
wishes store the i18n key alongside the translated literal, and the UI
renders the key when present.
The key is cleared server-side the moment `name` actually changes,
compared against the stored value rather than trusting the caller, so no
client can forget to clear it and silently revert a user's own name to a
preset label.
An unknown key (a stale row, a renamed key) falls back to the stored
literal via an isTranslationKey guard, so a bad key can never crash the
page or leak a raw key to the user. Rows predating this column keep
displaying their stored name.
_FULL_SCHEMA in db/connection.py is updated in step with the migration.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RXMhc9nT6sNz1SL4j8dznL
Works through every numbered item in NextSteps260913.md section 3, plus the
three known verification issues in section 4.1.
Surface what the backend already computed:
- 3.1 render surplus_advice in PlanResults, mirroring the shortfall remedies
- 3.2 surface VacationBalance.is_assumed so assumed statutory years are visible
- 3.3 age-4 expiry given the proposed plan, via a new Age4Projection reusing
the existing _age4_projections arithmetic (Dashboard stat verified, not rebuilt)
Calendar cluster (one pass over the holiday data):
- new GET /api/calendar/days exposing roda dagar, klamdagar and day overrides,
paginated four years at a time
- 3.4 shade roda dagar/klamdagar, with R/K badges so colour is not the only signal
- 3.6 click-to-override behind an explicit mode toggle, leaving drag multi-select intact
- 3.7 click month/year to advance
- 3.8 per-year red-days list in Settings, statutory days plus overrides
i18n:
- 3.5 five backend-generated Swedish surfaces converted to message_key + data,
following the tips_service pattern; old string fields replaced, not kept
Flexible wishes (3.9):
- PLANNING_ENGINE_SPEC section 10: window_mode, candidate search, cost function,
deterministic tie-break, two-pass ordering, test matrix rows 15-24
- backend: entity, migration 006, _FULL_SCHEMA sync, repository, allocator
- frontend: season-preset toggle in WishEditor, unplaced outcome rendering
Fixes (4.1):
- pin typecheck to the local tsc rather than npx resolving a newer TypeScript
- ignore generated coverage/ and dist/ in the ESLint flat config
- fix the flaky usePlanApply test properly, with waitFor rather than a retry
Tests: backend 361 -> 404, frontend 540 -> 704. Coverage 89.47% -> 96.52% lines.
Not yet verified in a browser.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RXMhc9nT6sNz1SL4j8dznL
jacobh2