Measured values with a point of slack, from scripts/coverage-thresholds.mjs:
atlas, audit, settings and unsplash all gained cases here.
Only those four. The script prints a block for every domain, and taking it whole
would have lowered src/nest/auth lines from 96 to 95 and tightened domains this
branch never touched, which would fail the open PRs for no reason.
unsplash is the one that needed the work rather than just recording it: dropping
the "any admin's key" fallback removed covered branches, so the parse-failure
ternary that was already uncovered started dominating the ratio and the gate
failed at 80.85% against a floor of 82%. Covered it instead of lowering the floor,
which put branches at 89.36%.
A parking booking spanning the trip showed up on every day, sometimes in between
planned activities. It now appears on its first and last day only, with drop-off
and pickup wording, and contributes nothing to the days in between, not even a
badge in the day header. Applied to the day plan, the mobile timeline, the PDF
export and the shared trip view, so a shared link shows what the app shows.
Rental cars stay as they are. The report names the car's smaller tag as the thing
it does not want for parking, not as something to change, and a car's middle days
move into the day header, which is what people expect from a vehicle they are
driving around.
The distance connector between legs is keyed off the row that gets hidden, so
hiding it naively takes the "2 km" label with it, which is what #1504 was about.
There is a test that fails without that guard.
Each segment of a stopover flight or train can also carry its own booking
reference now, since airlines often issue one per flight rather than one per
booking. Legs already live as a JSON array on the reservation, so this is an
optional field per entry: no migration, no contract change, and a booking without
it renders exactly as before. Readable everywhere it can be entered, desktop and
mobile, through the same blur treatment other booking codes get, plus the PDF and
the calendar description. create_transport and update_transport preserve it too,
otherwise they would accept the field and silently drop it.
Closes#1937Closes#1943
Interaction degraded the longer the map was used past zoom 5 and a reload fixed
it, which the report traced accurately: admin-1 geometry was fetched for every
country whose bounds crossed the viewport, the cache was never evicted, and the
whole accumulated set was torn down and rebuilt as one layer on every addition.
Panning past Canada or Russia added their geometry to every later rebuild for the
rest of the session.
Only the countries in view are rendered now, a signature over that set skips the
rebuild entirely while panning inside the same countries, and the cache holds a
dozen countries, never evicting one that is on screen or still loading.
One cause was not in the report and is probably the bigger one: a fresh renderer
was constructed on every rebuild. Leaflet registers a renderer as its own map
layer and removing the GeoJSON layer does not remove it, so the pane collected an
orphaned element per rebuild that kept updating on every pan for the rest of the
session, and at the country layer each orphan is a canvas of four times the
viewport area. Both renderers are created once per map now.
The region layer stays on SVG on purpose. Leaflet's stylesheet only makes SVG
paths mouse-transparent, so a canvas container would swallow the clicks meant for
the country layer below zoom 5.
Closes#1950
The role IN ('from','to') filter from #1486 is correct and still in place, but it
cannot see the case this issue actually reports. Joining AirTrail legs into one
booking with a layover stop is opt-in, so legs imported separately store the hub
as the legitimate `to` of one booking and the `from` of the next, which no role
filter can tell apart from two real arrivals.
Pair them where the countries are counted instead: an arrival and a departure at
the same airport, in different bookings of the same trip, within 24 h of each
other are a connection, and neither endpoint counts as reached. 24 h is the same
window the importer already uses to chain same-PNR legs. Doing it on the read side
means the databases that already hold these flights heal on the next request.
Only flights pair, and never cancelled ones, so a rental car picked up at the
airport still counts its country (#1366). A booking that returns to where the
chain started is a return flight, and that check compares distance rather than
airport identity, so flying home into a different airport of the same metro area
does not erase the country you spent the day in. A country with a place logged in
it, or marked by hand, is untouched.
Closes#1535
Searching worked for the admin and 403'd for every other member. getMapsKey read
the calling user's own row first and then fell through to "any admin" via LIMIT 1
with no ORDER BY, so the admin who pasted the key hit their own row while everyone
else got a non-deterministic stranger's. That fallback is what CLAUDE.md forbids.
The Places and Unsplash keys bill the install, not the account that pasted them
in, so they become instance configuration: an encrypted app_settings row, resolved
as env -> instance -> the caller's own row, the same shape #1772 chose for the LLM
endpoint. The per-user column stays as the last step because /me/api-keys has
always accepted a personal key.
The trailing migration carries an existing install's key into the instance row so
nothing degrades to OpenStreetMap on upgrade, but only where that key already was
the whole install's: as soon as a second row holds one, it is left alone, because
the instance row outranks everybody's own column and an upgrade must not move a
member onto someone else's key and bill.
Two things from the same report:
The key test button probed with field mask places.displayName while the search
asks for places.rating and more, which is a separate SKU and can be restricted on
its own, so the button could pass while every search failed. It now probes the
resolved key with the search's own mask.
Saving API keys wrote no audit line. It now writes one settings.api_keys_update
row naming which fields changed, never their values, comparing cleartext because
the random IV makes every ciphertext look different.
Closes#1939
A tag belongs to a user, not to a trip, and the place create/update body is an
open record the Zod pipe does not police — so any tag id could be attached, and
the tag projection ships tags.user_id, which answers who owns a tag the caller
has no way to see. collection_place_tags had it verbatim.
Scoped to the roster rather than to the caller on purpose: a member tags a shared
place, another member re-saves it and, because the update is delete-then-insert,
sends that id back. Caller-scoping would drop a co-traveller's tag on every
foreign edit; roster-scoping keeps it. Collections resolve the same question
through their own member list.
No client sends a tags array today, so this only closes the hand-crafted path.
knownUserIds asked whether a user exists at all, not whether they are on this
trip, so any id on the instance could be made a payer or a split member. Both
read-backs join users, and the settlement projection carries the avatar through
avatarUrl, which passes an absolute URL straight out — for an OIDC account that
is the picture claim.
Payers and split members drop off-roster ids, as before, only against the right
set. Settlements cannot: they name exactly two parties and both columns are NOT
NULL, so dropping one would either break the constraint or move money to the
wrong person. Both wrappers refuse the write and return null, which lands on the
404 the update path already had. That answer is deliberately the same one a
missing settlement gets, so the endpoint says nothing about whether an id exists
— and it retires the foreign key error that used to surface as a 500 and drew
exactly that distinction.
createSettlement grows the 404 its sibling had; the MCP tool reuses the error
string already at hand. Money rules are untouched: a settlement is a transfer, so
refusing one cannot unbalance the sum, and the cent arithmetic is unchanged.
setParticipants took the ids on trust. The caller is proven to be on the trip and
the assignment is proven to sit on it, but nothing said the same about the ids in
the body, and the read-back joins users — so a stranger's name and avatar came
back to the writer and went out over the assignment:participants broadcast to
everyone else in the room.
Takes the tripId both callers already hold and filters against the roster. Same
drop-rather-than-reject rule as bag members and reservation travellers: the
participants box posts the whole list on every edit, so a trip whose membership
changed underneath it would otherwise be stuck.
ASG-SVC-028 covered a foreign key error rolling the replace back. The filter
takes an unknown id out before the insert can reach it, so that path is gone and
the case now pins what happens instead: a list of nothing but strangers clears,
like the empty list ASG-SVC-018 already pins.
Holding packing_edit or todo_edit says you may edit this trip's lists, not that
any user id you name belongs to it. Both updateCategoryAssignees loops wrote the
body's ids straight through, and both end by re-selecting through a JOIN on
users — so an id from anywhere on the instance came back with a display name and
an avatar attached, in the response and in the WebSocket fan-out to the room. The
packing item recipient writes had the same gap.
packing already had tripRosterIds for exactly this and used it only for bag
members; its doc comment said "assigned to a bag", which is probably why the
other three writes grew up without it. The query moves to DatabaseService next to
canAccessTrip, which asks the same question as a predicate, and the helper now
delegates. Every service that needs it already injects DatabaseService.
Off-roster ids drop rather than 400: a copied trip carries assignee ids over
before its members exist, and rejecting the whole request would stall the picker
there. Guests are unaffected — a guest is a trip_members row, so the roster
includes them, which is the point of #1362.
A string handed to bindTooltip becomes innerHTML. JourneyMap passed an entry
title and a track name through untouched; the GL renderer beside it escapes
both already, and JourneyPublicPage imports the Leaflet one directly. The atlas
bucket-list tooltip had the same gap.
Also pins the escape on MapViewGL's photo marker. Its Leaflet twin has
FE-COMP-MAPVIEW-074 for this, but the GL side was only ever exercised with
benign values, so escapeHtml could have been dropped there without a test
noticing.
image_url and website were bare strings in the contract, and the create/update
bodies are open records, so the Zod pipe never looked at either. Both leave the
database for something that treats them as a URL: the thumbnail is interpolated
into hand-built marker HTML, and the homepage is handed to window.open, where a
javascript: value would run in this origin rather than opening a page.
Both are allow-listed on the way in now, next to hexColorSchema and checked in
the handler the way validateRouteColor already is, on all three write surfaces
so REST, the plugin RPC and MCP agree on what they accept. The entity schemas
stay open: rows written before this still have to round-trip.
safeHttpUrl covers the read side for exactly those rows, the same split
safeHexColor uses, and the window.open calls it guards pick up
noopener,noreferrer while they are being touched.
AssignmentOpsController declared @RequirePermission('day_edit') on its
mutations but carried only JwtAuthGuard. TripAccessGuard is what reads that
decorator and what resolves :tripId, so on this controller the permission
markers were metadata nobody looked at and the trip in the URL was never
matched against the caller. Its sibling, DayAssignmentsController, has had the
pair since the migration.
The two participants handlers also went straight to the service with the id
from the path, unlike move/time/transport, which ask getAssignmentForTrip
first. They ask now too.
updatePositions has two branches. The legacy one updates
`WHERE id = ? AND trip_id = ?`; the per-day one, three lines below it, bound
the caller's ids straight into reservation_day_positions and never touched the
tripId it was handed. That table has no trip_id of its own and its two foreign
keys only require the ids to exist, so nothing further down caught a pair that
did not belong together.
The trip now comes from the join, so the row materialises only when the
reservation and the day agree on it. Doing that in the statement rather than as
a pre-check also turns a stale id into a quiet no-op instead of a foreign-key
error surfacing as a 500, and lets an omitted day_plan_position bind the way
the wire contract has always said it may.
reorder_reservations reports a day that is not on the trip, the way its sibling
tools in that file already do, and a migration clears rows whose reservation
and day disagree.
The edit/duplicate/archive/delete buttons sat at opacity 0 until the card was
hovered, over a white .18 fill. Two people in a row said they never found how
to unarchive or rename a trip: invisible at rest, and white-on-white over a
light cover even while hovering.
They stay on the card now at .92, the fill is a dark scrim so the white icons
carry against any photo, and the icons get a heavier stroke. The status pill
and the hero badge move to the same scrim so the cover overlays read as one
set. An archived card shows the restore icon rather than the archive one,
which is what mobile already did, and every button gains a title next to its
aria-label.
The admin page always opened on Users, whatever the address said. A support
reply, an onboarding mail or a bookmark could point at the page but not at the
panel it was about, so every one of them ended with "then click X on the left".
Read once for the initial value and written back on each change, so the address
bar keeps saying where the reader is. Replaced rather than pushed: eleven tabs
would otherwise fill the back button with the history of a single page.
On a centrally administered install /api/admin/oidc answers 403, and the admin
page asked anyway. The promise was caught, so nothing broke, but the request
still went out and still printed a red 403 in the console of every admin who
opened the page. Swallowing a rejection hides a failure from the code, not from
the person reading the console.
The nav-entry seam was the wrong shape for what actually attaches here. A
subscription surface is about the install, not about a trip: /admin is already
behind an admin check, already grouped, and already where somebody looks to
change how their instance works. A top-level nav entry put it beside My Trips
for a screen opened once a month by one person.
Two lines in AdminPage rather than an overlay replacing the file: the tab list
gets the extra entries appended per group, and the body renders whichever one is
active. Appended rather than inserted because PageSidebar needs a group's tabs
contiguous.
The route and nav-entry seams stay. They are the right answer for a surface that
genuinely is a page, and both remain empty here.
Hiding the nav entry is not access control: the path stays typeable, and a screen
about how the install is run should not open for somebody who was invited to one
trip. ProtectedRoute already knows how to refuse — the attachment point just was
not asking it to.
Optional, so a route that does not set it behaves exactly as before. It sits
next to adminOnly on the nav entry rather than being derived from it, because
the two answer different questions: who is offered the door, and who may walk
through it.
The attachment point hands the two nav bars a static list, so anything attached
there showed up for every user of the instance. What belongs there is the
operator's side of the install, and a trip's other travellers are on the instance
as guests of the person who runs it — an entry about how it is run is noise to
them.
`adminOnly` is optional and absent means "for everyone", so nothing that exists
today changes meaning. Both nav bars go through the same exported filter rather
than each writing the rule, since a rule written twice is a rule that will differ.
Still inert here: the list is empty, and the tests pin that an empty list stays
empty for either kind of user.
FE-PDC-019 asserts that the collapsed row shows one specific line, but the row
shows the current day in the place's timezone — and the fixture gives Saturday
different hours from the working week. So the test passed Monday to Friday and
took the build down every weekend, which reads as a broken feature rather than a
test that forgot to say which day it is standing on.
Pinned to a Wednesday, with the Saturday case kept as its own test rather than
thrown away: it is the one that proves the day lookup happens at all, and it is
the exact case that was failing.
* feat(config): put the operator-owned surfaces behind TREK_MANAGED
The switch landed last release with nothing behind it. This is what it turns
off, plus the four fixes that came out of walking every one of these surfaces.
Nothing here changes anything for an install that does not set the flag. That
is not a hope, it is what managed-inert.test.ts asserts: the mode stays off for
any unparseable value, no other variable can turn it on, every settings key
stays writable, and splitManagedKeys hands back the same object rather than a
copy so no write path can quietly diverge on self-hosted installs only.
WHAT THE MODE WITHHOLDS
Fourteen routes carry @ManagedForbidden: key validation, static MCP tokens, JWT
rotation, the OIDC admin pair, restore and auto-backup, plugin sideload and
reload, the Ollama model list and pull, and the SMTP test send. Each names the
reason it is not the instance admin's, and the reason is mandatory.
Eighteen settings keys are filtered instead of refused. SMTP, the OIDC toggles
and the WebAuthn pair share one handler with keys the admin does own, and the
llm_* family is reachable through per-user settings, the instance defaults and
the addon config alike, so a route-level refusal would be both too coarse and
incomplete. A locked key is skipped and reported as managed_keys; the rest of
the save goes through. Refusing the request would take the registration toggles
down with the SMTP block.
GET /api/auth/me/settings stops returning the three API keys in the clear. It
is the only endpoint that decrypts a stored credential back to the caller, which
is fine when the admin pasted it in and wrong when the operator supplied it.
Two multipart routes check in the handler rather than the guard: a guard throws
before the parser drains the body and the client gets an ECONNRESET instead of
the 403 (PROFILE-015). They still carry the marker so the boot gate sees them.
WHAT KEEPS IT FROM DRIFTING
validate-managed-routes.ts refuses to boot on a marking that is not on the
reviewed list, a list entry whose route lost its marking, or a marking with no
reason. The third case has no equivalent in the guard gate next door and is the
one that matters most: a blank reason destroys the only record of why a control
was taken away, and nothing recovers that later. It found two wrong entries in
my own list while I was writing this.
managed-keys.test.ts asserts that every key in ADMIN_SETTINGS_KEYS and
DEFAULTABLE_USER_SETTING_KEYS belongs to exactly one of two sets. A key added to
either list belongs to neither until somebody decides, and that is the failure
this prevents: the lists grow, the mode does not.
PLACES_API_BASE AND PLACES_API_KEY
The nine Places endpoints spell out googleapis.com, so an install that wants
those calls to leave through something of its own has no way to say so. One
variable, substituted where every call already funnels through. A second one
supplies the key, and when it is set the users table is not consulted at all —
a per-user key would route around whatever the endpoint counts, and those
columns are readable by any admin of the instance.
FOUR FIXES THAT ARE NOT ABOUT THE MODE
An admin setting somebody's password now bumps password_version and clears
their MCP tokens. An admin sets a password for one reason, and until now
verifyJwtAndLoadUser kept accepting every cookie the intruder already had: the
one action taken to lock them out was the one action that did not.
The four OIDC provider calls go through the SSRF guard. Three are worse than
they look — the token, userinfo and JWKS URLs come out of the discovery
response, so they are chosen by the other end, not by an admin. safeFetchLlm is
renamed safeFetchAdminConfigured and keeps its old name as an alias, because it
is the variant that allows loopback and LAN: a self-hoster with Authentik on
localhost:9000 must keep working. The token exchange follows no redirect at all,
since one would hand client_secret to a second host.
DELETE /api/admin/users/:id/mfa clears another account's TOTP. The passkey half
has existed since passkeys landed; the TOTP half never did, which left
"somebody on the trip lost their phone" with no answer short of an operator in
the database. Never for the caller themselves.
TREK_PLUGIN_PERMISSIONS=off refuses to boot when the install is centrally
administered. A warning is right for a machine its owner chose to trust; where
the person who set the flag and the people whose data is in the database are
not the same person, fail closed.
TEXTS
Four strings lose the phrase "self-hosted" because they describe somebody
else's software — an Ollama endpoint, an AirTrail instance, a service a plugin
talks to — and read better without it in both modes. The one string that
describes TREK itself keeps it, and the mode picks a second key instead, so
self-hosters lose nothing.
A grep gate in shared/ keeps the canonical locale honest, because the parity
check compares the file set and the top-level keys and never looks at a value:
re-wording a string is invisible to CI today, and so is putting it back.
* feat(maps): supply the map credential with the instance when centrally administered
MAPBOX_ACCESS_TOKEN joins PLACES_API_KEY as an operator-supplied value.
SettingsService injects it when the settings are read, and picks GL as the
provider if the reader is still on Leaflet, since a token is always present.
The key moves from the customer set to the locked set. It is the one locked key
whose value is public by design — a pk.* token has to reach the browser to be
useful — so this is not about secrecy: a per-user save landing on top of the
instance token would only break the map for that user. Both settings fields for
it go away with it.
* fix(auth): count the operator key when reporting whether Google features work
has_maps_key only looked at the users table, but getMapsKey prefers the
PLACES_API_KEY env over anything stored there. On an install that supplies the
key that way and has an empty column, every lookup worked while the client was
told there was no key and hid the search, the autocomplete and the place photos.
Found by walking the running instance rather than the code: the map came up,
the search box did not.
* feat(admin): hide the settings a centrally administered install cannot act on
A control that answers 403 is worse than no control: the AI section rendered the
refusal as "Request failed with status code 403", which tells an admin they broke
something rather than that the decision is not theirs.
So the surfaces go rather than being disabled in place:
- AI parsing and AirTrail leave the addon list, and the AI addon route now
refuses outright. The endpoint, the model and what a document costs come with
the instance, so there is no instance decision left — not even whether it runs.
- Immich and Synology Photos leave it too. Both connect to a server the reader
runs at home, and there is no route from here to one.
- The SMTP block, the admin webhook and ntfy targets, and the passkey RP ID and
origins follow the same rule: they configure something the operator owns, and
a wrong RP ID on a shared parent domain reaches past this instance entirely.
- The About tab goes. It is about the project — what TREK is, where to file a
bug, where to support it — and is aimed at the person running the software.
The version and the link to the source move to the sidebar footer, on both the
desktop page and the mobile twin, because AGPL 13 wants that offer wherever
people reach the software over a network.
- The update check stops asking GitHub and reports nothing available.
Separately, the version footers lose the "self-hosted" suffix everywhere, not
just here. It read as a label on the software rather than on the install, and
the version alone says what it needs to.
* feat(config): refuse to boot centrally administered without an at-rest key
Measured on a running instance rather than assumed: with ENCRYPTION_KEY unset,
the at-rest key falls back to a file in the data volume, and backupService
bundles that file into every archive it builds. The archive is downloadable by
any instance admin, so the encryption ships with its own key in the same zip.
The skip in backupService already handles this — it only omits the file when the
key came from the env — but nothing made sure the env actually carried one.
Provisioning sets it; forgetting it was silent.
validateEnvAtBoot now carries a second kind of check next to the schema: cross-
field rules that only apply with the flag set, for combinations where every value
is individually valid and the result is wrong. Empty without the flag, so an
install that never asked to be centrally administered cannot meet it.
* feat(client): add an attachment point for install-specific screens
Two empty exports on the client, an empty module on the server, and the four
places that read them. Nothing here adds a route, a nav entry or an endpoint —
the arrays are empty and the module contributes nothing, which managed-ext.test
pins so it stays that way.
Read unconditionally, never behind a flag. On the server a conditional import
would tie the module graph to an environment variable, and the graph is built
once per app while the flag is read live; on the client an empty array costs
nothing to spread, so every build walks the same path instead of one build
walking a path nobody tests.
Anything attached here follows the same rules as the rest of src/nest and
src/components: a real module with a controller and a provider, bodies through
the Zod pipe, routes under the global guards, tests.
* test(shared): move the wording gate out of src so shared typechecks again
CI installs shared with its own dependencies only (npm ci --workspace shared),
so there is no @types/node and the `node:fs` import inside src/ failed to
compile. The tests passed the whole time — vitest supplies the module — which is
why this only showed up in the typecheck step.
The walk moves to scripts/, next to i18n-parity.mjs, and the spec imports it the
same way that one does. It also runs on its own now
(`node scripts/i18n-selfhost-wording.mjs`), which is what you want when CI names
a line you have to go find.
* test(server): repair the call sites typecheck:tests caught
tests/ sits outside the build tsconfig, so none of this showed up in a green
test run — which is exactly the gap `npm run typecheck:tests` exists to close.
Three kinds, all mine:
- Seven `import type` paths for RuntimeEnvService carried one `../` too many.
Type-only imports are erased, so vitest never resolved them and nothing failed.
- VersionCheckJob and PluginsController grew a constructor parameter and eleven
hand-wired `new` calls kept passing the old count. The argument arrived as
undefined and no assertion happened to read it.
- ConditionContext grew `managed`, and the shared context object in the
conditions suite did not.
The condition also had no test of its own, so it has one now: both directions,
since the registry uses `is: false` to keep a notice away and a notice that only
belongs on a centrally administered install would use `is: true`.
Some installs are run by somebody other than the person who administers them.
The operator owns the credentials, the configuration and the upgrade schedule,
and the admin of the instance is a user of the result rather than the one who
set it up. Nothing in the app knows that today, so it keeps offering settings
whose values are not the admin's to change, and the honest answer to half of
them would be a 403 the UI never anticipated.
TREK_MANAGED says it once. It is a single boolean and the surfaces that care
read it themselves, so adding one never comes back to the config layer.
Runtime-toggled, in the same class as DEMO_MODE, rather than a registerAs
token: roughly sixty tests mutate env mid-lifetime, and a value frozen per
built app breaks all of them. `=== true` on the way in, so an unset, empty or
misspelled value means off — a managed install sets it deliberately, and a typo
must not quietly take settings away from an admin who never asked for it.
The guard keys on a decorator, not on a path. MfaPolicyGuard documents where
the two literal path lists it replaced ended up: they stopped being maintained,
and every endpoint added afterwards silently fell on the wrong side. A marker
on the handler cannot drift away from the handler. It carries a mandatory
reason, which is for the next reader of the source rather than for the client,
so the marks stay auditable.
It registers third, after the auth and MFA guards, so a stranger still gets the
401 those two would have given them instead of a 403 that would confirm the
route exists. And it checks the mode before it touches metadata, because it
runs on every request of every install and almost all of them are self-hosted.
One 403 body, shared by REST and MCP. DEMO_MODE grew two texts for the same
condition and neither can be changed now without hunting for the other.
The flag reaches the client through app-config so the UI can stop offering what
the server would refuse anyway, but it is deliberately not persisted the way
demo_mode is: two installs can share a browser profile, and a stale "this one
is managed" would take settings away on an install that never set it.
No route carries the marker yet. The per-domain marks follow. Unset, which is
every self-hosted install, nothing about this changes anything.
An instance without a Google key already runs on the OpenStreetMap stack: search
falls back to Nominatim, autocomplete does too, photos fall back to Wikimedia.
The two details lookups were the exception. Both threw a 400 with "Google Maps
API key not configured" the moment they were handed a Google place id.
A Google id has no OpenStreetMap equivalent, so without a key there really is
nothing to resolve. That is an empty result, not a caller mistake. The callers
already treat a null place as a miss and fall back to the text-search path, so
they were doing the right thing with the wrong signal, at the cost of an error
round-trip and a logged exception each time.
It bites whenever a place that was saved while a key existed is opened after the
key is gone, and on any instance that never had one but imported Google-sourced
places from a shared trip.
The declared return type said the place was non-nullable while the OSM branch
could already return null and the wire contract has always had it nullable. It
now says what it does.
Google charges autocomplete per request unless the requests carry a session
token. With one, the keystrokes of a single search and the details lookup that
ends it are charged together, which is also how a user experiences it: they
looked up one place.
TREK sent no token at all, so every debounced keystroke was its own billable
request. On a busy trip that is the largest line on the Places bill, and it is
paid by whoever owns the key, which for a self-hosted instance is the person
running it.
A session opens with the first suggestion request and closes when the user picks
a result. It has to close on the abandoned paths too: a token carried into the
next search would merge two searches into one billed session and scope the
suggestions wrongly. So the search box clearing it, the text-search fallback and
the error paths all end it.
The token is validated on the way in. The autocomplete body goes through the Zod
pipe, the details query does not, so the controller shape-checks it there.
Anything that does not match Google's URL-safe 36-character form is dropped
rather than forwarded, which falls back to per-request billing instead of
breaking a lookup.
A cached details hit never reaches Google and so never closes its session. That
is billing-neutral: an unclosed session is charged as a plain autocomplete
session, which is still cheaper than the per-request behaviour it replaces.
A packing item is Common, Personal or Shared with named recipients. listItems()
always applied that rule, but every path that touches one item resolved it by
trip id alone, so trip membership plus packing_edit was enough to reach an item
the member could not see. Update, delete, clone, the contributor routes and the
sharing route all shared the lookup, and the same service methods sit behind the
MCP tools and the plugin RPC, so all three transports carried the gap.
The update was a read as well: it answered with the enriched row, which hands
back the name, category, quantity, weight, the owner's username and the recipient
list of an item the caller was never shown.
Two flows laundered restricted items into places they do not belong. Saving a
trip as a template read every row in the trip, so a template captured the names
and categories of everyone's Personal items. Copying a trip did the same and
re-inserted without is_private/owner_id, so both fell back to the column defaults
and another member's Personal or Shared item reappeared in the copy as a Common
item. That path needs only trip_create and trip access, not even packing_edit.
There is now one visibility rule in the service and every single-item lookup goes
through it. The actor is threaded through the HTTP, MCP and plugin-RPC entry
points, and a missing actor denies rather than falling through unfiltered. An
item the actor may not see is reported as missing, not as forbidden, so the
routes cannot be used to confirm that an id exists.
Two deliberate contract changes fall out of that. The sharing route answers 404
instead of 403 when the caller cannot see the item at all; a visible item they do
not own still answers 403. And a template or a trip copy now carries the Common
list plus the actor's own items, with a restricted item staying restricted and
owned by the copier.
Reported privately by Chichi (@vickyzer027-hash).
AirTrail 3.12.0 renamed the flight's passenger list from `seats` to
`passengers` and moved flightReason from the flight onto each passenger. There
is no alias in either direction: a 3.12.0 instance returns `passengers` and
`null` for `seats`, and its save schema accepts only `passengers`.
TREK read `raw.seats` everywhere, so against a 3.12.0 instance it found nothing.
The seat never made it into the import, and because a pull rebuilds the metadata
from the AirTrail payload, flights that had already been imported lost the seat
they had. That is the report. Flights whose sync is switched off were never
re-pulled, which is why theirs survived.
Two more breaks came from the same rename and were not reported. The flight
reason stopped importing for the same reason. And the writeback pushed the seat
under `seats`, which the new schema strips, so a seat edited in TREK was quietly
dropped on the way out; it did not fail loudly because the spread of the fetched
flight carried AirTrail's own `passengers` through untouched.
Both shapes are handled rather than switched between, because instances on 3.11
and older are still out there: reads take `passengers ?? seats`, and the save
body carries the manifest under both keys. Neither schema is strict, so each
version keeps the key it knows and strips the other.
The snapshot hash deliberately keeps feeding the manifest in under the same key.
An unchanged flight on an older instance therefore hashes exactly as before and
does not re-sync, while a 3.12.0 flight hashes differently once and pulls the
missing seat back in. There is a pinned-hash test guarding the first half.
Verified against real instances rather than the docs: a 3.11.1 and a 3.12.0
container, seat imported from each, a seat edited in TREK written back and read
out of AirTrail again, and the pre-fix build failing the same checks on 3.12.0.
Closes#1931
A stay covers a range rather than a moment, so one week in a hotel would hold a
slot in a six-entry widget for the whole week and push out the bookings that
actually happen on a day. The accommodation belongs to the day plan, which shows
it across its span.
They never worked here either, which is what the report is about. A hotel keeps
its dates on the linked stay, not on the reservation, so the two ways one can
exist behaved differently:
The booking form writes reservation_time = NULL and sends no day_id, so a
form-created hotel matched neither arm of the date test and was silently absent.
That is the "incomplete" half.
The accommodation panel auto-creates its booking with day_id and with the start
day's date stamped into reservation_time. That row did show, but updateAccommodation
syncs only metadata and the confirmation number to its linked bookings, so moving
the stay left the stamp untouched and the widget kept the old date. That is the
"no effect when I change the date" half.
The type test is NULL-safe: the column defaults to 'other' but nothing stops a
NULL, and a bare `type != 'hotel'` evaluates to NULL there and would have dropped
those rows from the widget instead.
Closes#1934
The empty state went in without flex-1 while the message list it stands in for
has it, so it was only as tall as its own content. The composer rode up to sit
directly under the mascot with the rest of the panel left blank, which reads as a
layout that failed rather than a chat waiting for its first message.
At a 12px gap the pill's own shadow ran into the sidebar it sits beside, so the
two read as one surface rather than a control floating on the map. 20px.
PDF, ICS and GPX each had their own button in the day-plan toolbar, and two of
them carried a hover menu of their own. That is five controls' worth of width in
a pane the user can drag narrower, so the exports were the first thing pushed off
the edge, and a hover menu is not reachable at all once its button is gone.
One button opens a dialog instead: the day plan as a PDF, the bookings as a
calendar (a one-off .ics or a live subscription) and the map data as GPX in the
three scopes a device wants. Every row reports that it is working, which the PDF
export needed most: it fetches a 226 kB builder before it renders anything, and
until now the toolbar button simply sat there.
Two strings the old ICS menu had hardcoded in English, the download and the
subscription, come from the locale files the phone export sheet already used.
The toolbar no longer needs the icsHover flag it only held to keep its menu open,
so that comes out of the sidebar's state as well.
The status, the time and the flight or train number sat inside one pill strung
together on a middle dot, and the dot was baked into the translation itself:
planner.resPending ended with " · " in all 23 locales. That made the row read as
one run-on sentence, and it meant the separator could never move without a
locale change.
They are separate chips now, same tint so they still belong together and nowrap
so the time cannot break away from the status it qualifies. The separator is out
of the locale files, which also retires the regex the transport detail modal
needed to strip it back off.
The chip that opens the label manager was labelled with collections.linkLabel,
which is the caption of a link in the list editor. English calls that one "Label"
so it happened to read correctly there, but German got "Bezeichnung", Russian
"Название" and Ukrainian "Підпис": wrong word, and long enough to push the filter
row out of its panel. It uses collections.labels.title now, which is the manager
it opens and is shorter in every locale.
collections.select was the other one over the edge. It is only an aria-label on
the desktop filter bar, so the eight locales whose word did not fit are on a
shorter one now.
The truncate on the chip labels never worked: a flex item keeps min-width:auto,
so the row grew past the panel instead of shrinking its chips. With min-w-0 a
long label name or a bumped-up text size ellipsises inside the panel.
The centre "+" matched /collections exactly, so it only knew it was on the
collections screen while no list was selected. Picking one moves the route to
/collections/:id, the match failed, and the action fell through to the last
branch of useCreateAction, which creates a trip on the dashboard. That is the
state the screen is normally used in, so for most people the button never added
a place at all. It also explains the report that there is no way to add one: the
workaround, adding from "All saved" and assigning afterwards, is the only path
that happened to still work.
The handoff carries the id rather than dropping back to /collections. Otherwise
adding from inside a list would land on the union and the sheet would ask for the
list the user is already looking at.
Closes#1930
Add CoMaps, the offline OpenStreetMap navigator, alongside the existing
map integrations. A geolocated place gains a CoMaps entry in its
navigation menu, and a day's route toolbar gains a button that hands the
day's bookended stops to CoMaps in that day's travel mode.
CoMaps' route link carries only a start and a destination, so a two-stop
day goes as a route and a longer one as named pins rather than silently
losing the stops in between. Links use the https form so a device without
CoMaps installed lands on comaps.at instead of a dead scheme.
The hotel-bookend logic behind the Google Maps export was duplicated
between desktop and mobile; both now share dayExportStops so the two
links cannot disagree about which hotel legs are real.
Addresses discussion #1904
fmtDateTime's time-only branch appended '00' for seconds no matter what it was
given. The booking import stores what KItinerary hands it, which is '10:00:00',
so the value came out 17 characters long. dtLine only attaches the TZID to a
well-formed 15-character value, so those events lost their zone and went out as
floating times, which is the #1453 regression by another route.
It pads the way the ISO branch above it already does, so every caller is covered
rather than only the rental markers.
Rentals now emit two additional timed markers, Pickup and Drop-off, the
same additive shape hotels already get for check-in/check-out (#1721,
last unported piece). The rental's existing single-block event is
untouched.
Sides are derived in priority order: the endpoint role (from = pickup,
to = drop-off, reusing the already-loaded endpoints map); if no
endpoint on the reservation carries either role, positional first/last
by sequence (only once there's more than one endpoint, so a lone
endpoint is never both sides); otherwise reservation_time /
reservation_end_time, since imports frequently carry only that window.
Per side, an event only fires once a usable date and clock resolve;
zone resolution prefers the endpoint's stored timezone, then its
coordinates, then the reservation's linked place.
Follows the hotel `marker` closure's pattern closely (same esc/uid/
dtLine/isDate/isTime/usedZones usage, same Pickup:/Drop-off: summary
convention, same per-kind UID shape) without touching the hotel path
itself.
The variables only take effect when Node runs with NODE_USE_ENV_PROXY=1, and
that is set in the image. This page also covers Helm, Unraid and source
installs, where nothing sets it, so following it there changed nothing.
Also recommend NO_PROXY. Without it the proxy sees every request TREK makes to
itself, including the container health check.
Review of the previous commit surfaced that the departure-time lookup keyed a
trip-wide coordinate index, so a POI revisited on another day (a supported
pattern) leaked the other day's time into the prefill. Resolve the departure
time from the leg's own day instead; the name index stays trip-wide (a name is
coordinate-stable).
Also from review: wire the per-leg option on the mobile sidebar too (it was
desktop-only), drop the now-dead route-leg fallback in the outgoing menu opener,
and accept single-digit hours in the time normalisation. Adds a regression test
for the revisited-stop case and pins the hotel-bookend fallback time.
The per-segment travel-mode menu now carries a "Public transit" entry on every
editable connector: stop-to-stop legs and the boundary hotel/booking bookends
alike. Choosing it opens the existing transit search pre-filled with the leg's
origin, destination and the stop's departure time, instead of drawing an OSRM
road route.
A connector only renders when both endpoints are located, so the option is
inherently limited to legs that can actually be searched. A leg's coordinates
are resolved back to named, timed points from the trip's places, accommodations
and booking endpoints. TransitSearchPanel gains an initialTime prop so the
departure time can be seeded.
Dependabot cannot be pointed at dev. \ arget-branch\ in dependabot.yml applies to
version updates only, and security updates always open against the default branch.
Security updates are all this repository gets, since there is no dependabot.yml at
all, so every one of them arrives on main.
The branch check then labelled them wrong-base-branch and close-stale-wrong-branch
shut them 24 hours later. Thirteen of the seventeen Dependabot pull requests died
that way; the four that survived were retargeted by hand. The comment the check
leaves asks the author to click Edit, which a bot cannot do.
So retarget them instead. If that fails, because the branch was cut from main and
dev has moved on, it falls through to the existing path so a human sees it rather
than the check quietly passing.
Note this only takes effect once it is on main: pull_request_target loads the
workflow from the pull request's base branch, and these arrive on main.
* fix(mcp): accept per-leg times on create_transport and update_transport
metadata was pinned to Record<string, string>, so a legs array could not get
through and a multi-leg flight built over MCP lost every intermediate time. The
booking ended up with correct endpoints and no legs, which every reader then
treats as a single segment.
legs is now a real, typed parameter with its contract in @trek/shared, shaped
exactly the way the planner form and the importers write metadata.legs. The
folding checks what the readers depend on: one leg fewer than endpoints (the
AirTrail dedupe builds its signatures on that), day ids that belong to the trip,
airport codes that agree with their endpoint, and the times, where a stopover
endpoint carries the departure and the last one the arrival.
metadata keeps its type and its replace semantics. Segments smuggled through
metadata.legs as a JSON string used to be stored and then ignored by everything
that read them; they are refused now and point at the legs parameter instead.
Closes#1914
* fix(atlas): refuse a bucket list entry that is already on the list
Nothing gave a bucket entry an identity, so the same country could be added as
often as it was clicked. A wish is now identified by name, country, target date
and, where it has them, coordinates. A different target date is a different wish
and still lands, which is what the report asked for; two real places sharing a
name stay separate because their coordinates differ.
Empty string and NULL are normalised on both sides of the comparison, otherwise
the dateless case, which is the one that was reported, would slip through.
Deliberately no unique index and no migration. better-sqlite3 is synchronous and
the server is a single process, so the race an index would guard against cannot
happen here, and an index would have needed a destructive dedupe of existing
rows first.
create_bucket_list_item had no target_date parameter at all, so over MCP the
allowed "same place, other date" case was unreachable. It takes one now. The
plugin RPC reports the duplicate as bad params instead of letting it surface as
an internal error.
Closes#1898
* fix(settings): keep a free-form LLM endpoint in admin hands
Any account could name the address the server calls, and the guard in front of
it deliberately allows loopback and LAN so a self-hosted Ollama works. That is a
reasonable trade for whoever runs the instance, because the network is theirs. It
is not one for an arbitrary account, which could aim a request from inside that
network at a target only it chose and read how the target answered.
A free base URL, and the local provider that is one by another name, now come
only from an admin controlled source: the addon config, the instance-wide
defaults, or an admin's own settings. OpenAI and Anthropic stay available to
everyone, because they go to a fixed address with the user's own key.
The resolver is where this is enforced, since it is the single point both the
booking import and the plugin host pass through, and the only one that also
covers values already sitting in the database. The write path refuses as well,
so nothing is stored that would never be read.
The hint text promised something users no longer have, so it changed with it.
Closes#1772
* fix(calendar): span a date-only booking to its end date
A DATE event without DTEND is exactly one day per RFC 5545, so a multi-day
booking that carries an end date still collapsed onto its first day. It gets an
exclusive DTEND now, built the same way the stay branch already builds one. A
booking without an end date serialises byte for byte as before.
While in here: the stays query joins reservations on the accommodation, and two
bookings pointed at the same stay fanned it out into two events sharing one UID.
Duplicate UIDs in a feed are undefined behaviour and calendar clients pick one at
random.
The same "at most one booking per stay" assumption sat in the accommodations
service, where it was worse than cosmetic: accommodation_id carries no foreign
key, so deleting a stay took only the first booking and left the second pointing
at something that no longer existed. Both the sync and the delete walk all of
them now. The callers keep reading linkedReservationId; the list is additional.
Refs #1869
* fix(vacay): say that the region is required, because it is
The placeholder called the region optional while the add button stayed disabled
until one was picked. The wording predates multiple calendars per plan, which is
what introduced the gate. Only the value of the existing key changed, in all
locales, so the key set stays identical.
Two things the same screen got wrong were in the way of a clean answer:
The button was clickable while the region list was still loading, which put the
user in exactly the state the wording describes, and it is the state the report
was filed from. It waits for the list now.
Changing the country on an existing calendar kept the previous country's regions
on screen, so a German region could be picked for the Netherlands. The list is
cleared with the country, which only works in combination with the loading gate:
without it, the empty list would happen to enable the button instead.
Closes#1813
* fix(collections): let a saved place's address be corrected
The edit form offered name, category and labels, and the address only on the way
in. The cause was not the form: the update contract never declared address, so
the validation pipe stripped it and the UPDATE had no column for it. Coordinates
were pulled through for #1435 and the address was missed.
It travels the whole chain now, contract to service to both forms. No migration,
the column has always been there.
The mobile add sheet had no address input at all, showing whatever the search
result carried as read-only, so editing would have worked while creating still
would not. It has one.
The MCP tool description had drifted further than the address, never having
mentioned the coordinates either. Both are in it now.
Closes#1870
* fix(collections): keep the place search usable after the first save
The add dialog stays open on purpose so several places can be added in a row, but
after a save it gave no way back into the search. A search with no results was
silent, with no empty state and no toast, so typing another name and pressing
enter looked like nothing happened. The reporter described both halves and it is
the silent search that matches "nothing happens" most closely.
There is an empty state now, using a key that already existed in every locale and
was unused. The search row is sticky inside the dialog body, so it stays reachable
after the save instead of sitting above the scroll position.
Also fixed while in here: refreshing the collection after each add reset the
select mode, the selection and the label filter. In the very flow this dialog is
built for, the user quietly lost all three.
Closes#1921
* fix(vacay): make the phone year grid and the dock button do what they look like
Two separate faults met on this screen.
Tapping a day in the year grid did nothing unless the edit mode was already on,
and the only way into it was a floating button that looks exactly like the one
the dock draws in the same spot. A tap now opens that month in the editor, where
the cells are big enough to hit. At 390px the grid cells are around 21px, which
is too small to be mutating data with in the first place.
The dock's plus was not wired for this route, so it fell through to creating a
trip on the dashboard. Two different actions were stacked in one place, and the
wrong one was on top whenever the screen had not painted yet, which includes the
time its chunk is still loading. The dock leaves the slot to the screen now, and
holds the geometry so the tab groups stay symmetric.
Closes#1811
* fix(budget): settle in whole cents so balances stop drifting
The report is a balance that moves on its own, with nothing added, edited or
removed, on a trip where everything is in euro. Two mechanisms were behind it and
both are real.
The display currency is a fourth value, a per-user setting, so "everything is in
euro" never covered it. Each balance was converted and rounded on its own, and
the set no longer summed to zero: at 0.855 and 0.9312 it came out one cent over,
at 0.94 it came out even. That cent appears and disappears with the six hour rate
cache alone, which is exactly the symptom. The flows were rounded separately from
the balances, so the two never had to agree.
And a receipt line nobody was ticked for counted toward the total, and therefore
toward the payer's credit, while landing in no share. That difference is
permanent, and the client sent it that way.
The settlement converts once, at the edge, and does integer arithmetic after
that. Equal splits distribute the credited cents by largest remainder rather than
the entered total, so credits and debits of one expense cancel exactly even in a
foreign currency. Display conversion goes through the same largest-remainder
allocation, so the sum survives it.
With that, the 0.01 epsilon can go. It could never create a cent, but it stranded
them: a one cent balance was filtered out of the flows while staying visible in
the list, which is the "settle up offers nothing" half of the report.
Closes#1382
* fix(journey): reach the book export and the suggestion toggle on a phone
Below 1024px the hero card is dropped, and it held the only way to either action.
The floating bar that replaces it was never given them. It has them now, and the
hero is unmounted rather than hidden, so there is no second invisible copy of the
same controls in the document. The bar's own buttons moved onto the appearance
tokens while they were being touched.
The phone screen never had either action at all and now carries both.
The book itself was hardwired to English, cover and dates included, while the
keys for it had been sitting in every locale unused. It takes the page's
translator and locale, sets the document language, and prints mood and weather,
which were on screen but never in the book.
Closes#1848
* fix(journey): offer the connected photo providers in the phone entry editor
The desktop editor grew an external photo source in #1632; the phone editor that
came out of the mobile rewrite never did, so editing an entry on a phone had no
way to Immich or Synology. It mirrors the desktop mechanics, including the part
that is easy to miss: a save that creates the entry and then fails on the photos
keeps the sheet open, and without remembering what was just created the retry
creates a second entry.
The sheet no longer kicks off its own addon load. The list is loaded once on
boot and this only reacts to it; a second load races the other consumers and
overwrites the store when it lands.
Closes#1808
* fix(mobile): let the document scroll so Safari can collapse its URL bar
iOS Safari only shrinks its address bar when the document scrolls, and TREK's
never did: the root was locked and the mobile shell put a scroll container inside
it. On the bottom tab layout that costs about a sixth of the viewport,
permanently.
Below 768px the document is the scroller again and the full-bleed screens measure
their own height instead of leaning on the shell.
That makes the body scroll locks real for the first time on a phone, and they
were not ready for it. Three of them reset the overflow unconditionally, one of
those on every run of its effect rather than on unmount, so a system notice could
hand the page back while a sheet was still open. They all go through one
reference counted lock now, which is the part that had to land first.
Worth knowing for the issue: this reaches the flow screens. The trip planner, the
journal, the atlas and the vacay screen keep their own full-height scrollers and
still will not collapse the bar.
Closes#1809
* fix(mobile): keep the screen gradient visible behind the translucent cards
The gradient layer introduced with the scroll change sat at z-index -10. A
negative layer paints above the canvas but below the backgrounds of ordinary
blocks, and body carries an opaque one, so the gradient was hidden behind it.
The mobile cards are translucent and take their depth from what is behind them,
so every surface ended up compositing against a single flat colour and the whole
phone UI read as one tone.
The layer is positioned at z-index 0 now and the content above it at 10, which
puts both over body's background. Same visual as before the scroll change, and
the shell still owns no scroll container.
* test(reservations): teach the accommodation delete stub about the id lists
The RPC walks every linked booking since the delete stopped stopping at the
first one. The stub still answered with the two singular fields, so the loop had
nothing to iterate and the cascade events never fired.
* fix(settings): move the free-form LLM endpoint out of user settings entirely
Hiding it from non-admins closed the hole, but it left the option sitting in the
admin's own personal settings, which is not what was asked for and not where it
belongs: an instance has exactly one address it calls out to, and there was a
second, invisible place to set it.
The provider list is now the two hosted ones for everybody, the endpoint field is
gone, and the resolver takes the address from the admin-controlled instance
defaults for every caller including an admin. It no longer reads the users table
at all. The write path refuses the keys for everyone, so nothing can be stored
that would never be read.
The endpoint is configured on the addon in the admin settings, which already
carries provider, model, base URL and key, plus the Ollama model list. Nothing
is lost, it just has one home now.
Behaviour change worth a release note: an admin who set a personal Ollama
endpoint has to move it to the addon config. A non-admin was already in that
position after the previous commit.
Refs #1772
* fix(journey): make the book preview usable on a phone
The preview document carried no viewport, so a phone laid it out at its default
980px while the page is A4 landscape and about 1123px wide. It spilled out of
the frame, which reads as being zoomed far in. Declaring the real width lets the
browser scale the page to fit. Printing is unaffected, it goes by @page.
The preview chrome had the same problem in miniature: a title and two labels on
one row, sized inline, which does not fit a phone and fits translated labels
even less. It moved to a stylesheet, and below 640px the card goes full-bleed,
the decorative title steps aside and the two actions share the width.
Refs #1848
* fix(mobile): stop the shell trapping the screens' own overlays
Lifting the background layer into the positioned range meant the content wrapper
had to be positioned too, and a positioned wrapper with a z-index is a stacking
context. Every fixed overlay a screen renders inside it got clamped to that
level: the vacay view/edit FAB sits at z-50 precisely so it beats the dock at
z-40, and it disappeared behind it.
The layer goes back to a negative z-index and the wrapper back to no positioning
at all. What it needed in the first place was for body to stop painting over it,
which is one line in the phone media block rather than a new stacking context.
* fix(mobile): let the phone background override actually win
A media query adds no specificity, so the themed body rule further down the file
kept winning on source order and painted over the shell's background layer. The
override is on html body now.
The gradient was back to being invisible, which is the flat single-tone phone UI
again.
* fix(vacay): draw today's ring around the day instead of replacing it
On the phone, today short-circuited the day-cell matrix and returned a ring on an
empty cell. A vacation day or a company holiday logged for today was stored and
counted in the totals, but the cell showed nothing, so it read as if the tap had
not taken.
Today is an annotation now, layered like the school band and the shared-calendar
rings, which is how the desktop month card has always treated it. The strong ink
stays only where the cell has no fill of its own, since --m-ink is light in dark
mode and would disappear on a pastel.
* fix(map): stop serving a stale GL style document from the service worker
The Workbox rule for Mapbox matched every URL under api.mapbox.com, and the
style document lives there too. Under StaleWhileRevalidate the map was built
from the previous revision on every load, with the refresh landing one load too
late, so a style republished in Studio with different label languages kept
rendering the old ones. The network tab showed a 200 throughout, because the
service worker's own answer is a real 200, which is why this looked like the app
rewriting text-field at runtime.
The style now has its own rule ahead of the tile rules, on NetworkFirst so it is
current whenever the network is, and still cached so the map opens offline. Its
own small cache too: sharing an LRU with three thousand tiles made the staleness
come and go at random. OpenFreeMap styles had the same problem and are covered
by the same rule.
Closes#1924
Maurice