Connecting an own server asked Ollama's /api/tags and gave up on any answer
that was not a 404, and every failure, a refused key included, read as "not
responding". The model list is now looked for at /api/tags, at the address as
entered and at /v1/models, an answer only counts when it is a list, and a
refused key or a server that answered without a list say so (#2808).
AI parsing through an own server always called Ollama's /api/chat, which an
OpenAI-compatible server such as Groq answers with 404. A server without it
now gets the same extraction over /chat/completions at the address as
entered (#2806).
The return address already carries a query, so a code appended with ? instead
of & ended up inside the ai marker and the page did nothing, without a word.
The server takes a sign-in only once. The clipboard read on return and a paste
of one's own could send the same address twice within one render, and an
address left over from an earlier attempt used up the one in progress, so the
real address came back as 'took too long'.
Every page of an install with the assistant and AI parsing off asked /api/ai/status and logged the 404 the addon guard answers. The store asks only while an addon opens the route and the install is not managed, and one hook on the desktop window and both phone screens asks again when the assistant or AI parsing is switched.
The demo account can look at its Account tab, API keys, MCP configuration,
own AI connection and language, but not change them, and the server already
refuses each of those writes. DemoLock renders such a section dimmed and out
of reach under diagonal stripes in theme tokens, with a small Demo mode badge
and one line saying why; useDemoAccountLock decides, on the desktop settings
and the phone screens alike (a new pair in feature-pairs.json).
Locked: the whole Account tab (profile, password, 2FA, passkeys, deletion),
API keys, the MCP card with its tokens and connected apps, connecting an own
AI model and the language picker, which shows English. The admin of a demo
instance sees everything unlocked, 2FA included, so the old instance wide
"Not available in demo mode" line under 2FA and its key are gone from all
locales.
MapLibre 6 and Mapbox GL 3 both throw out of their Map constructor without
WebGL2. Every place that builds a GL map now asks one helper first,
components/Map/glSupport.ts, and a browser without WebGL2 gets the Leaflet
map with the user's tiles instead of a map that fails to start:
- the trip map (MapViewAuto) and the journey maps (JourneyMapAuto) pick
their renderer through useGlRenderer, which folds in the Mapbox token
rule they each had;
- the settings preview, desktop and phone, shows the Leaflet preview
through useMapSettingsForm and says why under it;
- the tour planner keeps the 2D map when 3D is asked for and says that
3D needs WebGL2;
- the PDF route map and the Leaflet vector basemap already fell back and
now report a GPU failure too.
Behind the probe, buildGlMap wraps each constructor (MapViewGL,
JourneyMapGL, the settings preview, the terrain map, the PDF map). A GPU
that claims WebGL2 and still refuses the context is reported through
utils/webgl, which turns the answer to no and tells every mounted map,
so they switch to Leaflet without a reload; any other error is still
thrown. Two new strings, map.webgl2.fallback and map.webgl2.terrain, in
all 27 locales.
GHSA-jrc7-96c5-q579 covers every maplibre-gl up to 6.4.0 and has no 5.x
release, so npm audit kept flagging the patched 5.24.0. Moving to 6.15.0
retires the local backport: patch-maplibre.mjs and its postinstall,
predev and prebuild hooks go, along with the Dockerfile copy and its
format-baseline entry. 6.x sanitizes attributions against an allow list.
What the major changes for TREK:
- ESM only, no default export: the engine module imports the namespace
and hands that out as its default, so every caller stays as it was.
- The worker is no longer inlined, and inside a bundle MapLibre cannot
find it next to its own module. The engine sets it from Vite's
?worker&url build before any map, the Leaflet bridge's included.
- WebGL2 only: the probe asks for webgl2 alone, so a WebGL1-only browser
gets the raster basemap instead of a GPUInitializationError.
- 6.9 shapes right-to-left text itself and deprecated setRTLTextPlugin;
the old plugin would replace that and could only run through eval in
the module worker, so MapLibre no longer registers it. Mapbox does.
The lockfile moves maplibre-gl and its own dependencies only; the
Leaflet bridge already allows ^6.0.0 as its peer.
The card that opened claude.ai's custom connector dialog for this instance
goes, with its phone twin, the line in the assistant setup that jumped to it,
its endpoint and its contract. Anyone who wants Claude uses TREK's MCP
server under Integrations, or an Anthropic key in the assistant. It came with
the assistant and never shipped in a release.
- Tests that still knew the ChatGPT sign-in as personal only follow the
admin's shared plan: the service takes the caller with their role, the
admin tiles show ChatGPT, the controller tests cover the admin routes.
- The table refused an instance ChatGPT connection with a CHECK, so
sharing the plan failed. A new migration rebuilds ai_connections
without that check; the shipped migration stays as it was.
- useChatgptConnect grew past the function size limit: reading the
clipboard and noticing the return to the tab are helpers of their own.
- The file link factory hands its overrides over apart from the row,
like every other factory, which ends the deep type instantiation.
- Knip: exports only their own file used are private again, and the
booking cap reads MAX_WRITES_PER_STEP instead of an alias.
- lint:i18n-keys reaches the tools' own action names through a key
builder and the connection hints through literal keys.
- MMehrSheet stays in the assistant hints pair only.
The connection for everyone can now also be the admin's own ChatGPT sign-in.
Whether a plan may be shared that way is the self-hoster's call under their
plan and OpenAI's terms; the form says what it means before the sign-in.
The sign-in starts on the admin route and finishes on the shared one, which
refuses a sign-in for everyone to anyone who is not an admin. The ChatGPT
return reads the clipboard only once the paste handler exists.
- The sign-in opens in a small window over TREK instead of a new tab.
- Step 2 shows the address the window ends on, so the error page there is
expected rather than alarming.
- Back in TREK the copied address is read off the clipboard where the browser
allows it and the sign-in finishes; else the field has the cursor.
- The token and OAuth client factories hand their columns over apart from
the row, as the others do (TS2589).
OpenAI offers a fixed web redirect for Sign in with ChatGPT only to selected
commercial partners; open-source apps sign in through the 127.0.0.1 loopback
callback. The relay could never be switched on, so it goes again.
Once the TREK project has a client id registered with OpenAI, the sign-in
goes out with the relay page as its redirect address and the install's own
settings in its state; the relay sends the browser back there and the
settings finish the sign-in, nothing to paste. The relay sees the code but
cannot redeem it without the install's PKCE verifier. Without the project's
client id, or for an install without an https address, the loopback page
and the pasted address stay as they are.
relay/chatgpt-callback/index.html is the page to publish.
- A card for a tool without one of its own names what it works on: the place
of an assignment, the day, the note, the booking; its arguments sit under
the labels the other cards use, a new order reads as the names in order.
28 frequent tools carry their own action name in every language.
- The server's own steps without a card still refuse, as before.
- Tests: the rate limit is asked one request after another, the mention list
closes on the caret move a browser reports, the place factory hands its
category over as a plain id, eleven files formatted.
- agent loop and decisions without await in loops: steps as recursion,
calls and applied changes in order through one promise chain, the
independent reads and claims in parallel; part mapping in its own helpers
- place card, header and weather strip without refs or components created
in render; long hooks and components split; state reset during render
- trip notes only in a turn's tool session, never an undo's; map view
reported only while the assistant can run
- place strip scrolls through the strip's own smooth scroll, no physical side
- tests follow the raised limits and the new card fields; toasts read through
a mocked useToast instead of a window global; baselines lowered where the
counts fell
- the assistant keeps facts for a trip (remember_for_trip) as a change card,
applied into the user's trip notes; existing expenses can be changed
- change card fields as a few icon chips with the rest folded under Details,
the day band marks only what changes, the week grid as compact tiles
- no map button on receipts of things that have no place on the map
- tests follow the composer's accessible name, the receipt's status icon and
the GL map's container; the hanging stream helper honours an early abort
- Postgres probe: qualified columns in the upsert updates; entities restored
to the generator's output; strict ratchet reads the SDK client as untyped
The three AI addons share one tile in the integration column; a dialog holds
the connection, the assistant's limits and AI parsing, which now reads with
the assistant's connection unless an install keeps its own. Existing parsing
configs stay where they are and keep winning.
The chat sees the map on screen, offers next questions under an answer,
shows its live step in the window head, previews waiting changes inside the
day list, checks whether a day still works, folds a multi-day plan into a
week grid, keeps per-trip notes and can estimate costs. An empty or cut-off
answer gets one push instead of ending silently, and the assistant's own
step and size limits are raised.
Adds an in-app assistant that reads and plans the open trip. Users connect
their own provider (Sign in with ChatGPT, or an Anthropic, OpenAI, Gemini or
OpenRouter key) or use the one the admin set up, including a custom
OpenAI-compatible server. It is an addon and only shows while a connection
is usable.
The chat runs on the existing MCP tools, pinned to the trip. Every change is
held as a change card the user applies or discards, and applied changes can
be undone. On the desktop it is a floating window that can be moved and
resized; on the phone it is fullscreen. Search results show as place cards
that can go onto a day, weather lookups as a strip, proposed places as
ghost markers on the map with a before/after switch, and an answer can be
shared into the trip chat. Booking files can be dropped in and are read
into change cards.
Every help screen checked against the v5 redesign. Tours is documented for
the first time, as a screen of its own with twelve screens under it, and
notifications get help of their own. All pictures are taken again on a
realistic demo trip, with short loops where a still cannot show the motion.
The English is plain and every language is written naturally rather than
word for word.
The switcher shows two levels at most, a screen's overview lists its
sub-screens as cards above its guides, each wiki page is linked once, and
the help sidebar is wider.
Fixes found along the way: Save to Collection keeps a tour's route, a track
dropped on the places column opens the Tours import, tour Share on desktop,
the unsaved tour draft lives on the account, a route click no longer adds
an end point, Trip places leaves the trip's tours out, Escape closes a
stage's day menu, confirms before destructive deletes, DWD warnings sharing
one id kept apart, a booking's date moves with it, and smaller fixes across
notifications, bookings, the day plan, stays, Atlas, Collections, Vacay and
the admin screens.
Clears the client lint ratchet: helpers move to sibling .ts files for react-refresh, complex render paths get small helpers, floating promises are voided, the stage rail lets a stale request go during render, and the collections page reads the active list's tours and rights in ActiveCollectionTours. StageNumber is no longer exported.
The activity section used Footprints without importing it. Opening one stop's menu now closes any other. lostStageEnds no longer counts a stop the change kept without its mark as lost twice over. The wanderer watch switch keeps the same tree once the connection is known, so it is not remounted and keeps its focus.
The watch row only re-read its state when the connection came or went, so after
a passing connection test or a newly saved token it kept saying "Token refused"
until a reload. The connection card now hands it a key built from the saved
connection and its last check, and the row re-reads the watch whenever that
changes.
Without a connection the switch was only dimmed: the button stayed in the tab
order and was announced as active. ToggleSwitch gains a disabled prop that
disables the button and marks it aria-disabled, and the reason stays reachable
through a focusable wrapper carrying the tooltip.
Settings, Integrations, wanderer gains an optional background check of the
trails you imported, off until you switch it on. While on, TREK reads the
short record of each such trail at most once a day (never its GPX file) and
leaves a hint when the trail may have moved on: "may have changed" when only
wanderer's change time moved, "changed" when its name or description proves
it, and "not checkable" when the trail is gone, no longer yours or carries no
change time. Nothing is ever said to be unchanged. A new hint arrives as a
notification that opens the tour (a new ?place= link in the planner); the
tour's provenance shows it with the source check right beside it, and the
card lists all open hints in a dialog, each leading to its tour. The check
reads the source in full as before and spends the hint; taking a change in
stays the confirmed reimport. No tour is ever changed or deleted.
The check only asks through your own connection, only about trails your
account imported into trips you can still open, and stops when you switch it
off, remove the connection, wanderer refuses the token (until a test passes
or a new token is saved) or the admin switches wanderer off; hints and
snapshots stay. A run asks at most 25 connections for at most 10 records
each, one after another, and an instance that fails (an outage, a rate
limit) is left alone for an hour, doubling up to two days.
Server: migration 046120 adds the switch and the run bookkeeping to
wanderer_connections and the tour_source_hints table (one row per snapshot,
for its importer, gone with it). WandererWatchService and an hourly
wanderer-watch job on the CronRegistrar live in the tours domain and read
through WandererService; GET and PUT /api/integrations/wanderer/watch and GET
/api/trips/:tripId/tours/wanderer/sources/:placeId/hint (place_edit, like the
check) answer with @trek/shared contracts. A new user notification event,
wanderer_source_hint, can be muted per channel. Saving a new token now clears
what the last check said about the old one. There is no MCP tool, as for the
rest of the personal wanderer connection. All 27 locales carry the strings,
and the admin wiki page says what the check does.
- GPX files drop onto the map or a day; the import lands on that day
- wanderer connection per user in Integrations; the workspace offers the
user's own trails only, behind an admin switch under the Tours addon
- wanderer difficulty is stored on the SAC scale (T1 to T3), backfilled
- a GPX or wanderer import can be planned as an editable tour, either in
place (the recorded track is kept and can be restored) or as a new tour
that remembers the import it came from
- rails, rows and detail dialog follow the plan sidebar's cards and badges
- admin addon shelves with more than one row fold, collapsed by default
- reverse-geocoded waypoint names are cached on the device
API keys were split off from MCP to work without any addon, but the tab
that holds them still hid unless a photo provider, MCP, AirTrail, AI
Parsing or Dawarich was enabled. On a fresh install all of those are off,
so there was no way to create a key. The desktop tab now always shows;
the addon backed sections inside it keep their own gates. The phone layout
has no API Keys section and keeps its gate.
The wiki now says where keys are created and that no addon is needed.
Refs #2561
With the badge and toast options gone from usePhotoProviderConnections,
the desktop and phone cards called the hook identically and then derived
the same seven per card values (fields, config, values, badge, testing,
whether save and test have a route) line for line, which lint:dup counts
as a copied block of 25 lines in MPhotoProvidersSection.tsx.
The hook now hands each card those values through providerCard, and both
views keep only their markup. Nothing the user sees changes.
New test FE-COMP-PHOTOCONN-011 covers providerCard. No existing assertion
changed.
In the admin's default user settings, the map template dropdown listed
seven tile presets, while every user's own map settings listed nine: the
same seven plus Amap and Amap Satellite. An admin running an instance for
people in mainland China could not pick the one basemap that works well
there as the default, short of pasting its URL by hand.
The Amap presets (57eac32f9) were added to the two user preset copies and
not to the admin's separate copy; nothing in that change or later gives a
reason, and the server stores any template as a default. The later
refactor kept the gap as DEFAULT_MAP_PRESETS beside MAP_PRESETS. There is
now one MAP_PRESETS list, read by the user settings and the admin defaults
on both desktop and phone.
Regression test FE-ADMIN-DUS-029 (DefaultUserSettingsTab.test.tsx) was red
on the old code: "Unable to find role="button" and name "高德地图 (Amap)"".
Changed assertions:
- useMapSettingsForm.test.ts FE-COMP-MAPFORM-001: it expected the admin list
DEFAULT_MAP_PRESETS to hold 7 presets and the user list to add Amap on
top; it now expects the one list to hold 9 presets including both Amap
templates.
On the desktop, the Immich (or Synology) card could show "Connected" for a
server that was down. The settings page loads two answers at once: the
saved settings, whose connected flag for Immich only says that a URL and an
API key are stored, and the status route, which actually reaches the
server. Whichever landed last set the badge.
usePhotoProviderConnections had a statusRouteOwnsBadge option that only the
phone turned on. The option is gone: whenever a provider declares a status
route, only that route sets the badge, on both views. A provider without a
status route still takes the badge from its settings read.
Regression test FE-COMP-PHOTOPROVIDERS-028 (PhotoProvidersSection.test.tsx)
was red on the old code: "Unable to find an element with the text: Not
connected" (the late settings read had turned the badge to Connected).
Changed assertions:
- usePhotoProviderConnections.test.ts FE-COMP-PHOTOCONN-005: it expected the
later settings read (connected: false) to override the status route's
connected: true; it now checks that a provider without a status route
takes the badge from the settings read.
- usePhotoProviderConnections.test.ts: render() no longer takes the option,
so every call drops its true or false argument.
On the desktop, a failed photo provider connection test toasted "Could not
connect to Immich : Auth failed" with a stray space before the colon, and
"Could not connect to Immich " with a trailing space when the server gave
no reason. The phone already read "Could not connect to Immich: Auth
failed".
usePhotoProviderConnections took a testErrorGap option that put a space
between the sentence and ": reason", and the desktop passed ' '. The option
is gone, so both views join the reason directly.
Regression test FE-COMP-PHOTOPROVIDERS-027 (PhotoProvidersSection.test.tsx)
was red on the old code: "Unable to find an element with the text: Could
not connect to Immich: Auth failed" (the toast read "Could not connect to
Immich : Auth failed").
Changed assertions:
- usePhotoProviderConnections.test.ts FE-COMP-PHOTOCONN-010: the desktop
toast was 'memories.connectionError(Immich) : bad key', now
'memories.connectionError(Immich): bad key'; it also checks that a
failure without a reason toasts the bare sentence.
When the notification preferences failed to load, the matrix on both the
desktop tab and the phone screen stayed on its loading line for good and
nothing told the user anything went wrong. A failed load of the webhook and
ntfy settings was just as silent and left empty fields that looked like
nothing was saved.
Both loads in useNotificationSettings ended in .catch(() => {}) and had no
cancelled flag. Each failure now shows the common error toast, a failed
preferences load sets loadFailed so both views replace the loading line
with the error text, and both loads ignore a response that arrives after
the screen is gone.
Regression test FE-COMP-NOTIFSETTINGS-010 (useNotificationSettings.test.ts)
was red on the old code: "expected "vi.fn()" to be called 2 times, but got
0 times". FE-COMP-NOTIFSETTINGS-011 covers the cancelled flag and
FE-COMP-NOTIFICATIONS-032 the desktop view.
Changed assertions:
- MSettingsNotifications.test.tsx FE-MOB-SETNOTIF-002: a failed load used
to expect the "Loading..." line to stay; it now expects "Error" and no
loading line.
On the phone, every notification toggle sent the whole preference matrix as
it looked when the switch was tapped. With two toggles in flight the later
request could overwrite the earlier one on the server, and when the first
request failed the screen rolled that switch back while the second request
had already saved its value anyway.
The cause was the sendWholeMatrix option of useNotificationSettings, which
only the phone turned on. The server merges whatever cells it receives, so
the desktop already sent just the toggled cell. The option is gone and both
views send only the cell that changed.
Regression test FE-MOB-SETNOTIF-035 (MSettingsNotifications.test.tsx) was
red on the old code: "expected [ { trip_invite: { inapp: false,
webhook: false } }, { trip_invite: { inapp: false, webhook: true } } ] to
deeply equal [ { trip_invite: { inapp: false } }, { trip_invite:
{ webhook: true } } ]".
Changed assertions:
- useNotificationSettings.test.ts FE-COMP-NOTIFSETTINGS-004: the phone
toggle body was { trip_invite: { inapp: false, email: false } }, now
{ trip_invite: { inapp: false } } (the hook is built with the phone's
remaining option, skipMaskedToken).
- MSettingsNotifications.test.tsx FE-MOB-SETNOTIF-026: the PUT body was
{ trip_invite: { inapp: false, webhook: false } }, now
{ trip_invite: { inapp: false } }.
- MSettingsNotifications.test.tsx FE-MOB-SETNOTIF-032: the mock told the
in-app write apart by body.trip_invite.webhook === false, which only the
whole matrix carried; it now checks for the inapp key.
The live preview, the debounced save with its flush on leave, the colour mode save and the widget toggles live in one useAppearanceEditor hook. The contrast helpers and widget groups move to appearanceModel.
The tile presets, the provider and style rules and the preview place move to mapSettingsModel, the form state and save to one useMapSettingsForm hook. The phone keeps showing the normalized style before the save lands, through an option.
The preference matrix, the webhook and ntfy credentials and the channel test sends now live in one useNotificationSettings hook. The phone keeps sending the whole matrix on a toggle and skipping a token typed as the mask, through two options.
Profile and avatar, the password change, two factor setup with its backup codes and the account deletion now live in one useAccountSettings hook. Two existing differences stay as options: the toast key for a failed avatar removal and whether the backup code actions skip an empty code list.
Listing, adding with the step up and the WebAuthn ceremony, renaming and
deleting passkeys now live in one usePasskeys hook together with the
date and abort helpers. The desktop section and the phone card keep
their own markup.
The addon flags, the managed install flag and the addon reload of the integrations tab now come from one useIntegrationGates hook, and useSettings hands both shells the plugins tab flag next to the integrations one.
The OAuth presets, the endpoint snippets, the OAuth clients, sessions and API tokens with their handlers and the copy tick now live in one useMcpIntegration hook. Two existing differences stay as options: the desktop loads only while the addon is on, and the phone reads the client list back after a create or rotate where the desktop appends locally.
The declared field form with its save, the actions and their confirm state, the OAuth connect and disconnect and the activity log now live in usePluginUserSettings, usePluginOAuth and usePluginActivity. Each shell keeps its own form, confirm dialog or sheet and picker.
The provider field model, the seeding and hydration, save and test now live in one usePhotoProviderConnections hook. Two existing differences stay as options: on the phone the status route alone sets the badge, and the desktop keeps its space before the server reason in a failed test toast.
The provider, model, key and multimodal state, the hydration from the settings store and the save with its reload on failure now live in one useLlmConnection hook. The phone keeps its picker sheet state.
Both settings shells now read their state, the save and test handlers and the canSave rule from one useAirTrailSettings hook and keep only their own markup.
The count-string gate only knew {count} and {n}, the params t() picks a
plural form by. A quantity named anything else kept one form in every
language and passed: Russian read "Все {days} дней" for 2 days and
"{objects} объектов" for 3 objects, Greek "{minutes} λεπτά" for 1 minute.
Twelve such en strings become plural groups with their forms in all 27
locales, the quantity renamed to {count}: the trip length limit, the
chat image limit, the API key limit, the storage move prompt and its
reclaimable line, the Dawarich wish rule and suggestion duration, the
road trip lateness warnings, and the road trip echo and window texts.
Their callers pass the count as a number; the storage panels used to
pass it as a string, which no form is picked by. The reclaimable line
loses its dash for a comma in every locale while it is rewritten.
i18n-parity now treats the quantity names in QUANTITY_PARAMS ({days},
{minutes}, {objects}, {max} and the like) like {count}. The four strings
where the number is followed by nothing but a unit symbol (the Dawarich
durations, the upload size limit) go on NOT_PLURAL with that reason.
login, completeMfaLogin and register start tripSyncManager.syncAll()
and the system notice fetch without waiting for them, and every
sign-in, loadUser included, registers the sync triggers with their
30 s heartbeat. #2613 stubbed the sync in the push spec after it was
caught logging while the worker shut down ("Closing rpc while
onUserConsoleLog was pending"); the same shape was left in the other
specs that sign in for real.
authStore.test.ts ran the real sync on every successful sign-in and
logged DatabaseClosedError from it, and its notice fetch went past MSW
to the network and logged a failed fetch, both after the tests that
started them. The sync is now stubbed there, and the login, MFA and
register tests assert that it was started. The notice store is reset
before each test, since FE-STORE-AUTH-024 left a stubbed fetch behind.
LoginPage.test.tsx signs in through the form, and AccountTab.test.tsx
and MSettingsAccount.test.tsx reload the user after an MFA change; they
now stub the sync triggers (and the login spec the trip sync), which
none of them is about.
GET /api/system-notices/active gets a default MSW handler returning no
notices, so a login or a mounted notice host anywhere in the suite
settles that fetch in the test instead of on the network. The notice
store's own tests still override it.
Keeping archived trips stopped the sync from wiping them, but left nothing
that ever took one off the device. The trip sync never refreshes an archived
trip, and the 7-day rule and the per-trip switch only looked at the active
list, so an archived trip stayed until Clear cache however long ago it ended.
The stale and switched-off rules now also run over the archived trips that
are stored here, with the same eviction as an active trip: its rows go, its
parked changes stay. The offline settings list the stored trips the server's
list leaves out, so an archived trip gets a switch the user can turn off;
before, the switches came from the active list only and could not reach it.
Arabic runs the app right to left, but margins, paddings, insets, text
alignment, borders and corner radii were written for the left side, so
icons sat against the wrong edge of their labels, indents and badges
pointed the wrong way and amount columns aligned left. They now use the
logical forms (ms-, ps-, start-, text-start, marginInlineStart,
insetInlineEnd and their CSS equivalents), which are the very same sides
in a left to right language. Geometry stays physical on purpose: maps,
positions measured from the pointer or a bounding box, time axes,
sliders, transforms and print output.
Amounts drawn as several spans keep their order in a right to left row,
the time picker keeps hours before minutes, and two switches whose thumb
slid out of its track in Arabic now move the right way.
lint:rtl counts what is still physical per file against a baseline that
may only shrink (1512 uses in 370 files before, 319 in 113 now), and CI
runs it, so new code reaches for the logical forms.
Maurice