Vela Maps roadmap¶
What is still open. FEATURES.md lists what the app does and
SPEC.md how it is built. What shipped, and what was tried and does not work, is in
docs/ROADMAP-HISTORY.md.
Add an idea here when it comes up. When an item ships or is proven dead, move it to the history file in the same commit.
North star¶
A Google Maps replacement that needs no Google account, key or Play services, matches Google
Maps feature for feature, and over time depends less on Google by growing Vela's own data.
GPLv3. Every new data flow is opt-in and documented in PRIVACY.md.
The order from here¶
Set 2026-10-07.
- Finish and verify what is in flight. The week's bug reports, then a check of everything on the main phone.
- Android Auto without Google Play. The next big push. A car lists only navigation apps that Play installed. The next section has the details.
- The Compose toolkit upgrade. core-ktx 1.19, navigation-compose 2.10, hilt-navigation-compose 1.4 and the newer material3 pull Compose from 1.7 to 1.10 or 1.11 across the whole app. It waits until it can be its own job: a full pass over the sheets, the gestures and D-pad operation, on both test phones.
- Read places from Vela Almanac, a separate open dataset of US places with dated evidence of what is still open. It replaces the three closure steps in the places bake and later the merge itself. Blocked on its first published files. The bake then reads one file per state.
- F-Droid's own catalog. The official catalog builds every app from source and cannot take the prebuilt voice runtime (sherpa-onnx) or the prebuilt Cronet. The work is a build flavor that compiles or leaves out each of them, and reproducible output. Not started.
Traffic incidents from official feeds (issue #688; Germany's Autobahn API, Mobilithek and the Hamburg traffic feed in issue #833) are out of this list. They belong in a sister project beside Vela Almanac, which Vela would read as data.
Less to Google¶
The direction for the Google half: keep what only Google has, and send it less. Candidates, none built:
- Traffic re-checks from a point ahead. A drive asks Google for a fresh route from its live
position about every two minutes (
NavSession.maybeRecheck). Asking from a junction a kilometer or two down the route gives the same traffic for the road ahead without the live position. The time to that junction is the phone's own estimate, and an offered faster route has to be joined to the stretch still being driven. - Routes computed on the phone with Google's traffic. The on-phone router already answers offline. Fed congestion read from the traffic overlay's tiles, it could answer online too, and Google would see which map tiles were loaded and never a start, an end or a position. Open questions: what the tiles encode per road, how far out they must be fetched for a long trip, and how the times compare with Google's own.
-
A first run that starts from less. Someone who picks "Use Google" gets every default at once, the two-minute re-check included. The same holds for someone who starts without Google and turns it on later.
-
Google by choice on the place sheet. A tapped place asks Google at once, and the switches that hold it back are in Settings ("Look up tapped places on Google", "Load reviews only when I tap", "Load photos only when I tap"). The sheet could open on what the open data knows (name, address, phone, OpenStreetMap hours) with one clear control to load Google's hours, reviews and photos, so asking Google is something the user does and sees. It is the change reviewers ask for most (see "What reviews and forums keep raising").
- A request count that explains itself. Settings, Privacy shows how many requests went to Google in seven days. A light afternoon of testing reads as more than a thousand, with nothing to say that one place sheet or one drive is many requests. Show it per action ("a place: about N", "a drive: about N an hour"), and use the same numbers to find what to cut.
Android Auto without Google Play¶
What is known:
- Android Auto lists a navigation app only when Play installed it. Its "Unknown sources" switch does not cover navigation apps.
- What the check reads is not settled. One car log shows Android Auto asking Play for each
app's owners and denying a package that has none. A later report says the record that
counts is who started the install (
initiatingPackageName). - Nothing run from adb changes who started the install. On Android 14
-isets the installing package and the initiating package stays the shell. - It works with a root installer that runs the install as Play, and on stock Android with an installer that goes through Google's package installer, which GrapheneOS does not ship. Aftermarket head units with their own receiver list Vela after King Installer plus an ADB install.
- An in-app update replaces the install record, so on a phone set up for the car the updater
offers the APK as a file (
InstallSource.setForCar). - The Desktop Head Unit skips the install check. It tests the car screens only.
To do, in order:
- The ownership experiment.
-PappId=<id>builds Vela under the id of an app the phone's Play account once installed. If the car lists that sideload, the check is Play's library record alone. If not, the signing certificate counts too. Needs a real car and a stock phone signed in to Play. - Settle on the pinned issue (#179) which install record the car reads.
- Run the instrument-cluster session (
ClusterSession) in a car with a cluster display. No car has run it. The Desktop Head Unit's cluster option forwards turn data only, and that works. - Check the car map gestures (pan, pinch, fling) on a head unit. The Desktop Head Unit takes taps only.
- Test one Android "AI box" dongle (Carlinkit, Ottocast and similar): a small Android device on the car's USB port that shows its own screen on the head unit, with no install check. Wireless Android Auto adapters do not help. They still run the phone's Android Auto app.
A phone-side sender is the other route: an app that speaks the Android Auto protocol to a
real head unit in place of Google's app, so no install check is in the loop. The open
implementations (aasdk, openauto, the Rust android-auto crate) are all the head unit side.
The sender is Gearslip (Seb3thehacker), a separate project. It can be distributed only if
the head unit accepts a certificate we can generate. Key material taken from Google's app or
from head unit firmware cannot be shipped.
Vela's side of that is a small bound service that hands a projection client frames at a size
it names, takes input back, and publishes the nav state. CarMapRenderer already renders to
a bitmap and NavSession publishes what ManeuverMapper reads. Whether to hand out frames
or a URL for a WebView is open. Nav state and input are the same for both.
What reviews and forums keep raising¶
The same points come back in every review and thread about Vela, in several languages. They are kept here so a decision about wording, defaults or a feature can be checked against them. Where one has turned into work, the item is named.
- It is a copy, not an alternative. Vela keeps the Google Maps habit (the star rating picks the restaurant, Google's traffic picks the road) and removes only the account. The dependence is made comfortable, not cut. Position: this is what Vela is for. It is a way off the Google Maps app for someone who would otherwise keep it, and the north star is to need Google less over time, not to pretend it is not there. Apps built on open data alone take the other road, and for anyone they already serve they are the better choice.
- Asking Google should be a visible choice. Reviewers who accept the trade still object that it happens on a tap, without the user deciding. Before Vela they opened Google Maps in a browser, on purpose, and knew they had. Open item: "Google by choice on the place sheet" under Less to Google.
- It takes from Google's users and gives nothing back. Google's traffic exists because its users report their position; Vela reads the result and contributes none. True, and it cannot be fixed toward Google without an account. What Vela can give back goes to the open side: "Contributing back to OpenStreetMap" under Big bets. One user has already added their local transit agencies to Transitous because Vela showed the gap.
- Google can turn it off. By changing its pages, by serving anonymous sessions less, or through its lawyers, given how closely Vela follows the look. The first is handled by the signed calibration file and the second is shown in the app when it happens. The answer to all three is the same: with Google off Vela is still a working maps app, and every item under "Less to Google" and "Vela's own record of what is still open" makes that version better.
- "Degoogled" promises more than the app does. A packet capture shows Google hosts, so the
word gets called false. The README and site now say what is not on the phone (Google code, an
account, a key) and call Vela a hybrid of open maps and Google. Still to change: the welcome
tagline and the About hint in the app, and the F-Droid repository description.
PRIVACY.mdshould also list every host a capture will show, including the ones Google's own pages load in the hidden WebViews. - It is written with AI. Stated in
SECURITY.md, and the templates ask which model helped. Reviewers who raise it mostly conclude the risk is bounded because there is no server, no account and no credential to lose. The automated checks inSECURITY.mdexist because this was asked. - One person maintains it. Raised as a durability risk, fairly. "Opt-in telemetry" and "A Google Play listing" under Big bets are the two things that would change what one person can support.
Things reviewers got stuck on, each small:
- Street View could not be found on a place (#724). Fixed: the place sheet's photo strip starts with the Street View tile.
- A typed search returned a different place from the one autocomplete offered for the same text.
- Android Auto was reported as missing, which in a car is true today: see "Android Auto without Google Play" above.
In progress¶
Nothing on a branch at the moment (2026-10-12). Contributor pull requests are read in full and built here before they land; what each one still needs is under "Built, not yet checked on a drive".
Next up¶
Small items, one pull request each.
- Grid cells for the whole catalog. The bake conductor runs the US group and two world shards
(
tools/bake-schedule.json). Next: confirm fromcells-manifest.jsonthat every region has cells. - A name index for the places archive. Offline search reads the archive in rings out to
about 3 km (
PlacesArchiveSearch.MAX_RINGS). Beyond that only the OSM place pack answers, and OSM lacks whole chains. Bake a per-region sidecar (name, key, category, lat, lng) and query it after the pack. - Region downloads pull the building overlay. A saved map area downloads the Microsoft
footprints (
downloadOverlayForArea) and a whole-region download does not, so a downloaded state has no houses where OSM is thin. No bake change. - More places sources for the open bake: chamber-of-commerce member lists and municipal business-license registers, one scraper per source. The survey under "Vela's own record of what is still open" says how thin those registers are.
- A Takeout CSV import runs as a plain coroutine: no Cancel, no foreground service, and the
list is written only at the end, so a 2,000-row file is a quarter hour the process must
survive. Run it through
downloadLaunchwith a Cancel on the status card. - The blind-travel hold short of a turn (#745) holds the arrow, but the engine still advances the step and says the turn line from the held point once the reckoned speed is over 6 m/s.
- A stop whose pin finds no mark on the line pro-rates its countdown over two legs.
- The car map does not follow a map look change while connected; the look loaded with the surface stays until the next size change. Offline with a look that was never cached, the car gets no style at all, where the phone falls back to the bundled one. A session that started offline never picks up the patched style until the process restarts.
- The first-line shortening measures speech in characters, so a Japanese, Chinese or Hebrew line is read as short and keeps its sign tail.
- Duplicate names in the open places bake (#742). Overture carries one unbranded name at several points with the same confidence, so a pin lands on the wrong one. Two fixes at the bake: one row per unbranded name in a region, and a corrections list Vela maintains.
- The neural voice's phonemizer. espeak, in front of the Piper model, misreads text that is
not prose: "5:49 PM" came out as "five foot nine".
SpeechTextworks around it case by case. The fix is a model that normalizes its own input and is fast enough on a Pixel 4a. - Why one basemap layer stalls a dense city. With Liberty's
poi_r20layer drawn, New York at 100 ft runs at 0 to 25 fps on a Pixel 4a, and at 59 with it hidden. Vela hides it over region archives of rev 20260923 or later (placesOneSetRev). Elsewhere the stall remains. The cause is unknown. - Offline timetables per region, an open question. Cached boards cover only stops opened online. A bake of GTFS stop times is tens of MB for a mid-size state, plus feed discovery per agency and a weekly refresh.
- Review topic chips on the inline Reviews tab. Google's "mentioned in reviews" chips show
only on the full-screen reviews page (
PanelControlsinPlaceSheet.kt). - "People also search for" after an address or list open. It arrives only with a focused search, so it needs a focused name lookup.
- Menu tab reliability. Classify gallery walks that return no tabs. Find out whether the
gallery request (
hspqX) can filter by category, so the Menu tab needs no page walk. - Lanes that continue into the next maneuver, highlighted on a compound maneuver. OSRM gives no lane linkage across steps, so it needs a careful heuristic.
- State and province shield shapes from the OpenStreetMap Americana set.
ui/map/RoadShieldsdraws one plain badge for every state route. - A "download this region" nudge when an avoid option is on, Google cannot be reached and no downloaded region covers the trip.
- Parking: a note or photo on the saved spot.
- A resumed drive keeps its stops. After the app is killed mid-drive, the resume offer restores
the destination, label and mode only: the remaining stops and a saved route's hidden points
are gone (
NavController.persistNav). - On-street bike lanes. Painted lanes (
cycleway=lane) are not in the OpenMapTiles schema and need a baked layer, never a per-viewport Overpass query. The painted-roads test under "Richer roads" draws them. - Street View: walking can step to a neighbor from a different year (the neighbor graph carries no date), higher-zoom tiles on pinch-in, and hiding the pill where there is no coverage.
- Map label font. To match the app font exactly, rebuild the glyph pack from the app's font
file (
scripts/build-map-fonts.sh) and republishmap-fonts. MapLibre cannot inherit a font at runtime. - D-pad hardware pass on a real keypad phone: pan step, OK-hold threshold, focus ring
visibility, traversal order, page scroll on the full-screen reviews page. Device tests 04 to
06 in
dpad_test_suite/count key presses for one screen size and need target-seeking navigation, as tests 01 and 02 have. - Voice library. Host the catalog (
PiperCatalog) on the signedcalibration.jsonso new voices need no app release, with the download host pinned in the allowlist. A preview button. One sharedespeak-ng-datafolder (about 10 MB a voice). Larger dictation models. - Japanese offline voice. Piper has no Japanese phonemizer, so Japanese guidance uses the system voice. Kokoro int8 multi-lang (about 126 MB) needs the multi-file sherpa plumbing restored, and it ran at about 0.4x realtime when last bundled, so measure first.
- Explore: a sheet of nearby restaurants and things to do from the bare map. Events have no keyless source.
- Timed speed limits from a downloaded region. The hosted tiles apply
maxspeed:conditional; the obf path reads OsmAnd's plain limit, so a downloaded region shows 100 where the tiles show 130 at night. Needs the obf bake to keep the tag andcurrentRoadLimitto read it.
Built, not yet checked on a drive¶
Each of these passed its unit tests and has not been seen or heard where it matters. Take one off the list when it has been.
- Highway steps on a route named from the tiles (
LineNamer) come from the link roads: a turn straight onto or off a highway at an intersection says "Turn left onto", an exit is said once where the line leaves the highway and the end of the ramp is a plain turn, curves along a ramp or a highway say nothing, and two turns onto one street within 50 m are one. Checked with the tile step census on the 39 captured Google lines and the Davis I 80 route against Google Maps' steps; not yet heard on a drive. -
Exits on those routes take Google's exit number, signs and side ("Take exit 4A toward I-5 N: Redding"), and an exit the tiles miss is said where Google has it. Checked on 7 trips kept with Google's replies (16 exits, all said once); not yet heard on a drive, and the phone's reply has not been looked at for the same step fields (only the Mac's).
-
Unnamed car parks show their P under the open places layer and Google's dots (issue #935,
vela-osm-parking): seen on the 4a at the reporter's lot; frame rate measured in Shinjuku. Not looked at in a city with many unnamed lots at z15 to z16, where clutter would show. - Temporarily closed places (issue #930): the "Reopen this place" entry in a place's edit menu marks the closure when Google sends no status text. Checked on the 4a against open San Francisco listings (no kind 3) and a permanently closed Davis one (kind 3). No temporarily closed listing was at hand to see its sheet.
- A phone's own text-to-speech engine (issues #949, #944): its prompts are labeled navigation
guidance, and it keeps the voice picked in its own settings when that voice speaks the
directions' language. Neither test phone nor the emulator has a system engine, so neither has
been heard yet: check with
adb shell dumpsys audio(the player's usage) and an engine with several voices a language. -
Fewer voice lines on the freeway: any maneuver at 50 mph or more, and an exit, fork, keep, U-turn or on-ramp turn from 34 mph, gets one line about 15 s out and one at the turn. A merge is said once, often as the "then" of the step before it, and a "then" carries the lanes. A recorded drive replayed through the engine went from 24 lines to 17, and an exit, the fork on its ramp and the merge from eight to four. A Davis mock drive on the 4a said an I 80 on-ramp, its merge, exit 75 and the turn after it as expected. Not yet heard on a real drive: whether a quarter mile is enough warning to change lanes.
-
A roundabout is said on its approach and only buzzes at its entry, unless the entry line carries the next step: replayed on a recorded drive, the roundabout got its two approach lines and none at the entry. Not heard on a drive, nor through roundabouts in a row.
- A fork said as the "then" of the fork 45 m before it is not said again at its own turn
(
CHAIN_REPEAT_S): unit-tested at 25 m/s. The trip replay passes no speed, so it cannot show this. Not heard on a freeway interchange. - A reroute whose first step is a ramp, exit or fork the car cannot reach in 5 s asks again past
it (
PLACED_STEPS): unit-tested. Not seen on a freeway reroute. - The line at a ramp, fork or keep names the first route on its sign, and a fork toward the
road the car is on says "to stay on" it (
repeatWithRoute,stayOnForks): unit-tested and replayed on a recorded interchange. Not heard on a drive; "stay on" not seen on a real route. - A reroute turns down routes that turn the car round (back along its own track, backing up,
or set off against it) for one that goes on and costs little more, then for one that turns
round further on over one that turns round where the car is, from both routers or from one
alone when the open router is out; asks again 3 s later when the route turns a moving car
round where it is and nothing better was offered; keeps the last course at a stop unless the
track turned off it; and waits 3 s, not 10, after a route the car never joined. Unit-tested,
and re-driven on a recorded Davis drive with two missed turns on the 4a: the old code rerouted
four times and sat 30 s on a turn-around; the new code rerouted once per missed turn, set
Google's turn-around aside both times, and its route matched the road driven. The track
starts at the car's last turn-round: on a mock drive on a divided Davis road, the reroute
after the car's own U-turn took its route at once, where the uncut track held it 7 s. With the
open router off (
openRouterDown) on the Davis mock drive, Google's alternate that went on led at the first missed turn, and at the second, where Google's one route turned the car round, the re-ask fired and its answer went on (adopted after 4.2 s, no reroute after). Not seen on a real drive. The held-course turn check is unit-tested only. - A recorded drive keeps the nav engine's inputs (
N) and every router answer (RC), and a replay feeds those inputs back: checked on the 4a, where the replay went off route and swapped routes in the drive's order. Also on the 4a with a Davis mock drive: a seek to three quarters ran both reroutes silently and played on from 3:00 of 3:54 on the drive's last route; two 10 s GPS gaps left 10 dead-reckoned points, and the replay went through them with no false off-route; a trip recorded beforeNlines replayed its six reroutes in order. Re-drive (replayLive) used on the 4a only. Not tried on a real drive's file. -
A missed exit on I 80 at Davis (mock drive on the 4a): every answer went on to the next exit, none was flagged as turning round, and the reroute was adopted in 1.2 s; missing the next exit too, the second in 1.0 s.
-
Data job fixes: the camera bake asks for CSV and tries again, the maxspeed bake skips a region with no tagged roads, a failed velamaps.org deploy no longer stops GitHub Pages, and the baseline profile run warns instead of failing when it may not open a pull request. Each ran locally against the live servers and files. They are on main and none has run on GitHub yet.
- The car's time left is colored by traffic (#928): seen on a real Android Auto connection on a debug build. Not seen on a release build, on the desktop head unit or in heavy traffic.
- Android Automotive 10 and 11 keep Vela in the home screen's map pane (#879): on the Automotive 14 emulator the pane, the app list and Home behave as before, with one task, and on a 4a links and shares reach the running map and the "maps app" entry follows the "Maps" name. Not seen on an Automotive 10 or 11 car, which is what it is for. A share from inside another app now lands in Vela's own task; that was not tried by hand.
-
A ferry leg on a walk or a ride says "Take the ferry" (#824): seen in the step list on a 4a, walk and bike. Not heard on a real crossing. The phrase in the 17 other languages has not been read by a native speaker, and the on-device router's ferry legs are not phrased this way.
-
A line waits to play until it can play through (
StreamStart, #795): on a 4a with a high-quality voice at normal speed it starts as before. Not tried on a slow phone or at a raised voice speed, which is where the mid-line pauses were reported. - The voice engine's own threads take the worker's priority (#795): a test line under eight busy loops on a 4a rendered in 11 to 15 s instead of 25 s. Not heard on a hot phone on a drive, and the far prompts (speaking priority, nice 8) were not timed under load.
- A drive with Avoid on and no Google route takes Valhalla's avoiding route (#878): on the 4a with Google off, a Davis fixture trip moved off I-80 and the "may still use" note was gone. Not driven, and a trip with stops still gets no avoiding route without Google or a region.
- A shared run whose open route piece leaves Google's line by more than 15 m is drawn on Google's line (#895): unit test only. The reporter's trip today draws the same before and after on a 4a, so whether this was their loop is not known.
- A snapped route to a place ends where Google's line ends instead of at the pin: unit test. The one store tried on the 4a has Google's own end 25 m from the pin, so no route there changed; not yet seen moving a route.
- Offline, the downloaded region under the view's center is drawn instead of the one the view came from (#902): on the 4a, Montenegro then Kosovo downloaded, airplane mode, the center moved into Kosovo with Montenegro still in the west of the view, and Kosovo mounted. Not tried with regions of one country, and the neighbor beside the mounted region still draws blank.
- The phone arrow fitted to screens narrower than 1080 px (#911): builds. Not seen on a small screen.
- Takeout pins at 0,0 imported from their link, and a lists import of a saved places file pointing to the right import (#920, #921): unit-tested. Not tried with a real Takeout file.
- The free-drive follow turns heading-up only on speed held over 4 s (#905): builds. Not tried on a phone or on foot.
- A Google place asked for as the place, so the route ends at its entrance point
(
DestinationPlaces): checked live on twelve fixture stores, the built request answering the same route as the web client's own form. Not seen on a drive. - Full traces from the multi-select share (Save to phone, Share full traces): builds. Not tried on a phone.
- A recheck on the same roads never offered as a faster route (
RouteGeometry.sameRoad): unit-tested. Not seen on a drive. - A bare turn just past a ramp start said as a keep, and a lane clause on a ramp not saying the side twice: unit-tested, and a recorded loop ramp replayed says "Keep right". Not heard.
- Arrival short of the last turn only after 10 stopped fixes: unit-tested. Not seen on a drive.
- A fork on an exit ramp kept as its own step instead of folded into the exit: unit-tested with a fork 60 m and 300 m past the exit. Not heard on a drive.
- A bare exit worded "Take the ramp toward" the road after it, and a next step due at once said with the turn ("Take the ramp, then keep slight left"): unit-tested, and a recorded drive replayed through the voice went from 30 lines to 27 with none a second apart. Not heard on a drive; the replay re-folds the recorded steps rather than a fresh route.
- Street names filling in mid-drive (
RouteHeal). A simulated drive never re-checks, so it needs a real one. A trip's notes show it: a route whose totals line saysnamingLateorbare=1, thenrecheck upgraded route for names. - A route built past the naming deadline (
StretchNamer.inHand): tile names where the match was late, and a finished match kept. Seen on a phone only with the matcher switched off (debug.vela.tune.noMatch), never with a late one. - The nudged route line on a street. It was seen on a phone over a parking-lot stretch, where
Google's line was already near the middle. The totals line says
nudged=<meters>. - Two turns said on one line. Not heard on a phone. Dutch, Swedish and Japanese have no joining
line (
NavStrings.thenNext), and the joined line is not among the lines the voice renders ahead of time (NavEngine.upcomingPrompts). - The next stop on Android Auto, and the trip's own time beside it. Built, not run on the Desktop Head Unit.
- Side streets' stop signs left off a drive (#805): the bake's road for a sign from drivable ways and the most minor road at a junction, and the 10 m limit for a sign with no road. Needs a rebaked region driven down a main street, and one drive with no region (the Overpass path).
- The approach prompt that is not said twice three seconds apart on a short step, and the first line of a short step that drops its sign tail so the line at the turn does not cut it. From a 2026-10-11 trace; unit-tested, not yet heard on a drive.
- The car map's heading-up, tilted follow with no route, drive mode, the zoom by speed, the traffic raster and the phone's extra fills (#749, pull request #763). Seen on the Desktop Head Unit with a simulated drive and with a mock trace at 100 km/h and no route: heading-up, tilt 55, zoom 15.2 to 15.7. Snapshot time at that speed, median and 90th percentile: 37 and 51 ms with the traffic raster off, 40 and 65 ms with it on, five frames over 300 ms in 80 s against two; the raster is not what slows the car map. Not seen: a real car or head unit, and the rebuild of the car style when the patched file appears after a plain load, which needs a first launch with the car already connected.
- The project's links from one place (
VelaLinks) with a signedlinkspush to move them (SPEC 11.4). Unit-tested; no bundle withlinkshas been pushed yet, so a move has not been rehearsed. - Driving mode on the phone (#727): the drive's bar with an X, the locate and voice buttons, the speed box from the first fix, the search bar alone, the speeding alert from the fixes, and "Where the map opens" > "Driving mode". Seen on a 4a in portrait (the way in from the layers panel and from the setting, the voice choices, the X, the credit above the bar, a simulated drive started from it and ended back in it) and on a tablet emulator in both orientations. Not seen: the speeding alert, location off, the screen turning off, a real tablet, or Android Auto connected.
- Camera alerts in driving mode (#727): the look ahead along the fix's course. Seen on a 4a fed mock fixes at 34 mph toward a plate camera facing the road: the scan found it every 4 s, and the Heads up card and the spoken line came about 150 m before it, once. Not seen: a speed camera (needs a baked region on the phone), a camera past a bend, the car screen, or a real drive.
- Contrast map colors (
CONTRAST_DARK): seen on a 4a at zoom 12.6 and 15.5 on the Davis fixture, online, in the dark theme. Not seen: the car screen, a drive with the route line and traffic over the tan streets, the satellite layer, the AMOLED theme (which ignores it), the offline basemap, or terrain shading. - The map looks (#764): the offline fallback and the no-network guard were seen once, for Dark and Fiord online then offline over one region (Dark is now named Black).
- The lists sheet (#765): a physical keypad, and the sheet with one list, where a pencil stands
in for the handle. Open: the overflow button on a recent search row cannot be reached by key,
since it sits inside the focused row as on every
SuggestionRow, so a recent row has no key path to Remove from history. - Romanian guidance (#766): the phrases need a native speaker. After the merge the transit words
need a signed settings push:
stație,garăandautogarăintotransitCategoryWords,benzinărieandîncărcareintotransitExcludeWords, since a pushed list replaces the compiled one. The status words need no push. - Takeout lists (#767): a real Takeout export, and the import carrying on after Settings is left with its outcome on the status card.
- Timed speed limits from
maxspeed:conditional(#728, Malse-Makker): the parser is unit-tested on the Dutch tags and the badge still reads the streamed tiles on a simulated drive. Nothing changes on a phone until the maxspeed tiles are rebaked with the tag kept, so the 130 at night has not been seen. - Bridge openings in the Netherlands (#729, Malse-Makker): seen on a Pixel 4a against the live NDW feed on simulated drives, a planned opening's card with the right clock time, an open bridge's card and spoken line, and the chime in Alerts only. Not seen: a real drive.
- The three map buttons on the AMOLED theme (#769), every menu's corner (#771), a recent row's swipe to its menu (#772) and a renamed address keeping its street (#736): seen on a Pixel 4a. Not seen: Home and Work set from a plain address after the change.
- Square map buttons (#797): seen on a 4a, the locate and parking buttons with Round and with Square, and the setting reached from the settings search. Not seen: the keypad highlight on the square shape, the AMOLED theme, or the car screen.
- A reroute after a wrong turn on a trip with stops. The simulated drive cannot leave its route.
- UK fuel prices on a phone in the UK.
- Saved pins at every zoom (#791): seen on a 4a at zoom 7, where a pin 100 km from the fix drew at near full size. Not seen: the smaller pins below zoom 7 and a pile collapsing to the pins that fit, which needs saved places in more than one city.
- The layers panel as picture tiles (#782,
LayersPanel). Opened on a 4a in portrait (dark theme) and on a tablet emulator in landscape (light theme): it fits under the button without scrolling, and the tiles switch the look, satellite and traffic by touch. Not yet seen: the AMOLED theme, a keypad walk across the rows, and the panel on a shorter screen, where it would scroll. - "Zoom to the next turn" (#781): the zoom that frames the next maneuver under the turn card, unit-tested on the 1080 x 2340 fixture. Seen on a 4a in simulated drives: zoom 16 to 17.9 through town, out to 12 and flat once the next maneuver was 8.9 miles off on I 80, back in from 12.4 to 17.2 over the four minutes before an exit, and the turn 27 dp above its row while the camera held 52.6 degrees against a framing for 55 (fixed, SPEC 4.7 Framing), and after the fix the turn on its row, 24 dp under the card within 2 dp, through the tilted stretch. A driver's screenshot showed the callout half under the card at 24 dp; the margin is 72 dp. Not seen: a real road with it, a reroute with it on, and the floor dial below 12.
- A saved address or pin lists the businesses on it under "Also at this location" when it is reopened. Not yet seen on a phone: a saved address that a shop has moved into.
- Lights and stop signs as the map's own glyphs on the road-ahead bar (#828): seen on a 4a in a simulated drive through town, the light housings on the track beside a camera badge. Not seen: a stop sign's octagon on the bar, and the two 16 dp marks a few meters apart.
- Wide fixes and a phone at rest (#818, #819): a fix wider than 30 m under 3 m/s gives no course
and no speed, the car's heading holds while stopped instead of re-asking the route segment,
and the accelerometer says when the phone is at rest. Unit-tested (
StillnessTest). The car's route heading is read at the engine's progress along the route, as OsmAnd and Organic Maps do, and the segment nearest the fix only off the route; seen on the desktop head unit in a demo drive through town, the arrow along the route up to the turn. Not yet seen: a real stop at a light with a poor fix, on the phone or the head unit; a demo drive cannot make a wide fix. - Streamed overlays read by Vela instead of MapLibre (#827): the address, building and places
overlays of a region that is not downloaded come through
StreamedPmtiles. Seen on a 4a (2026-10-18): a Davis block at z18.6 drawing the overlay's house numbers, with its tiles asked ofpmtiles.vela.invalid. Not seen: footprints in a suburb OSM has not drawn, the open places layer with no region downloaded. - Compass words after a route number cut to their letter on screen (#837 follow-up): "toward
US 169 North" on the card and the car's step reads "toward US 169 N" (
RoadNameShort .routeCardinals,RoadNameShortTest). Seen on a 4a in a simulated drive: "Take the ramp on the right toward I 80 W". Not seen on the car. - The debug export's
inputline (#744): not seen; written for the head unit's report. - One list exported on its own (#822): a share button in the list editor's title row, seen on
a 4a: it opens the share sheet with
vela-list-<name>.json, and Delete, Cancel and Save keep their row. A file of that shape picked under Saved places > Your lists > Import came back as the list with its two places and its pins on the map. Not seen: the exported file itself going round, since the share sheet on the rig has no file target. - Names under saved pins from z14 (#823): seen on a 4a in dark and light mode at z14 to z16: three pins in a downtown grid with their names under them beside the open-places labels, a pin on top of the basemap's own label for the same place keeping the pin and losing the name, a nine-pin list across a city center with every name that had room and the crowded ones dropped, and names of six words wrapping to two or three lines at the 8 em width.
- The drive's
location:note (#818): providers on, and GrapheneOS's network location, SUPL, PSDS and geocoder choices. Not yet seen: a trip export carrying it; the GrapheneOS values need a GrapheneOS phone. - "Zoom out as far as" (#781): seen on a 4a, the three choices under the switch, Town picked by default, Region picked and kept. Not seen: a drive with Region or Far picked, where the map may go out to zoom 10 or 8.
- The camera turned toward the next turn, and Zoom to the next turn on the car screen (#781):
unit-tested (
TurnYawTest). Seen on a 4a in a simulated drive from Davis to the airport with the Town floor: heading-up through town with no yaw, the camera held 45 degrees toward the far junction through the looping on-ramp, then on the freeway the junction 21 km out kept in the center column (x 542 of 1080) with the camera 10 degrees off the road's heading, at 59 fps. Seen on the Desktop Head Unit (1280 x 720) in the same drive: zoom 12 and flat on the freeway with the camera 19 to 31 degrees toward the junction, back to zoom 15.2 and 55 degrees of tilt as it neared, the yaw fading from 43 to 23. Not seen: a real road, the yaw fading out through a turn in town at the speed zoom, and a head unit in portrait. - Whisper base and small for voice search (#798): seen on a 4a, both in the picker under Search with their sizes, base downloaded and unpacked in half a minute, made active, and the mic opened on it. Unit-tested: the first download offered is tiny for English and base for any other app language. Not seen: a spoken search through base or small, how German comes out, or the offer itself on a phone in another language.
- The lead-in before spoken lines (#802): 300 ms of silence at the start of every line, on Vela
Voice (
PiperSynth.LEAD_IN_MS) and on the phone's engine (playSilentUtterance, whose utterance the focus listener skips). Seen on a 4a: a simulated drive's prompts through the speaker, whole, four lines in a row with their timings in the log. Not seen: a Bluetooth head unit, where the clipping was reported, or the phone's own engine. - The turn card's Then line wraps to a second line (#799). Not seen on a drive here, which needs a compound turn onto a long street name; the change is the line count.
- Update all for downloaded regions (#801): seen on a 4a with four regions pending. The button showed the count, the regions updated in turn under one card, and a place pack whose delta did not fit fell back to the whole file as a row's own Update does. Not seen: Cancel mid-run, or a drive started during the run.
- The transit guidance pane's last card ends with Done (#811): seen on a 4a, a four-leg bus trip stepped to its end, the card's Back and Done, and Done back on the map.
- An expanded transit route stays expanded across a swipe of the sheet (#812,
TransitExpanded): seen on a 4a, a bus itinerary expanded, the sheet dropped to its card by its handle and raised again with the legs still open. - The transit guidance pane scrolls its legs, and Back and Next keep their height (#853). Not seen on a phone: a current walk leg with its steps open, a ride leg with its stops open, the list back at the top on each step, large fonts, landscape, and D-pad focus through the list.
- One Share on the route picker (#814): seen on a 4a, Start and Add stops below, Route options, Share and Close in the header.
- Saved routes named from a named start (#815): seen on a 4a, a trip from a named start saved under both names by default, and the list row with the start and the end under it; a route from the phone's own position still reads Route to the end.
- Drawings and custom maps above the satellite imagery (#816): seen on a 4a, a three-point line and its area drawn over the photo, and with the Vela look as before. Not seen: the satellite turned on with a saved custom map already showing, which goes through the same re-seat.
- Turn arrows on the route (#804): seen on a 4a in a simulated drive, a white arrow in a dark casing through the right turn onto a cross street and the turns after it, the passed one gone with its step, and after the corner was moved onto the line's own bend, the arrow through a freeway on-ramp bending left with the ramp and the blue turn bubble's tail on the corner; unit-tested on an L-shaped route, a keep-left whose split lies 60 m past the point and a lane change before a corner. Not seen on a phone: a fork found past its point, a lane change smoothed out, a roundabout exit, the switch under Navigation turning them off mid-drive, or the car screen, which has none.
- Turn arrows cut from the drawn line and redrawn inside it (round tail, plain head, thin indigo edge; reports of arrows beside the line and a head that read as stuck on): seen on a 4a in a simulated drive, a paused right turn with the arrow inside the line and no seam at the head; unit tests for the drawn line and a curve. Not seen: a Google route whose drawn line was moved to the road center, a tight curve on a ramp, or z18 on a phone.
- The car's speed badge and limit sign sized with the screen (#817): seen on the desktop head unit's 800 by 480 profile in a demo drive, the badge and the sign at the template's button size. Not seen: the 1280 by 720 profile with the new sizes, or a real head unit.
- The traffic layer's tiles stop at z16 (#800): seen on a 4a at z17.5 over a downtown, two z16 requests for the view where there were fourteen at z18, the bands scaled and even. Not seen: the reporter's phone, where close-up tiles never arrived.
- Each installed region's row ends with its newest bake's date (#801): seen on a 4a, two regions reading "updated Oct 1, 2026". Not seen: a region whose routing file is newer than its pack.
- The car's speed badge off by default (
CarSpeed, under Show speed and speed limit): seen on the desktop head unit's 800 by 480 profile in a demo drive, the limit sign alone in the corner. Not seen: the switch turned on, or a unit with the badge and the sign side by side after it. - The car's drive screen cut to mute and end up top, recenter and overview on the edge, and the route at 20 px in the phone's blue (user 2026-10-15, after Google's car screen side by side): seen on the desktop head unit at 1280 by 720 in a demo drive. Not seen: a real unit, or the landing and preview screens, which keep their zoom buttons.
- The camera's turn toward the next turn capped at 90 degrees instead of 45 (#781, a tester
found turns dragged off the side of the screen on curving roads): unit-tested
(
TurnYawTest). Not seen: a drive where the next turn lies more than 45 degrees off the heading with the yaw in full. - One accessibility list on a place page instead of two (#832): the chip row above the rating card is gone, the About sections below carry the attributes. Seen on a 4a: a grocery's page, its four accessibility items once, under About.
- A simulated drive from a chosen start (#835): Start drives the planned line from the start picked in the chooser; a real drive still re-plans from the car when the chosen start is farther than 150 m away. Seen on a 4a: a drive from a grocery 500 m from the dot began in its parking lot, on the 3.1 mi route the chooser showed.
- Road numbers first (#836, Settings > Navigation, off by default): a numbered road goes by its
number on the pill, the car's Continue on line and the directions of the next route. Unit-tested (
RoadLabelTest). Seen on a 4a: with the switch on, a simulated drive's card read "Turn left onto CR E6" for West Covell Boulevard with the number's chip beside it; off, a boulevard with a number reads by name. - One Save and one Share on a place page (#830): both are pills in the action row and the header keeps only the X, as Google lays it out; the Save pill opens the menu the header's bookmark had (quick save, lists, note, icon, name, Home, Work). Seen on a 4a: a grocery's page with the X alone up top, Directions, Start, Call, Save, Share, Website and Street View in the row, the Save menu (Save, Save to list, Set as Home, Set as Work) and the Share menu off their pills.
- Street words abbreviated everywhere on screen (#837): the pill for every name, the turn card
and its "then" line, the car's step and Continue on; the voice keeps the full words. Unit-tested (
RoadNameShortTest). Seen on a 4a in a simulated drive: "Turn left onto G St" on the then line and "G St" on the pill. - The car's speaker button stepping on, alerts only and off (#842): seen on the desktop head unit at 1280 by 720 in a simulated drive, the glyph going speaker, bell, crossed speaker and back to speaker. Not seen: the phone's own button following a change made in the car, or a real head unit.
- The phone's light and stop sign glyphs on the car map (#844):
CarControlGlyphTestfor the size; seen on the desktop head unit at 1280 by 720, the traffic lights through a downtown at about road width. Not seen: a stop sign on the car map, the 800 by 480 profile, or a real head unit. - The rating card's Show reviews loading the reviews when they wait for a tap (#845): seen on a 4a with "Load reviews only when I tap" on, one tap opening the Reviews tab with the reviews loading. Not seen: a place whose Reviews tab is already open when the card is tapped.
- Street View links opening the viewer (#847):
MapLinkParserTestfor the three link shapes; seen on a 4a, a web map Street View link handed to Vela opening the viewer on that block. Not seen: themap_action=panoandlayer=clinks on a phone, a link with a pano id, or the kav app itself. - A pair of close turns said in a line that ends before the second turn's own (user,
2026-10-17):
NavEngineChainTest. Not seen or heard: a real drive, or a trip export of the reported one, so whether the cut was this or the second turn's queued approach line is open. - Offline address search (user, 2026-10-17): the pack's street ids read as 64-bit, and the
word lists (
OfflineStreetWordsTest). Seen on a 4a withdebug.vela.tune.offline 1against the Northern California pack: a Davis address typed plain, with its town after a comma, with town and ZIP and no comma, and half typed ("Covell Bl") all led with the address and the store on it; before the id fix all four found no address. Not seen: a phone with its network really off, or the user's own area. - The Street View thumbnail for every Google listing (user, 2026-10-18): seen on a 4a, the Davis Food Co-op, which had none, now shows one; four other businesses kept theirs. Not seen: a listing whose nearest pano is across a wide road or behind the building.
- Save to phone for a recorded trip (user, 2026-10-18): seen on a 4a, the system save dialog
opening in Download with the trip's file name, and the full trace (15.9 KB, every line kind)
landing there for
adb pull. Not seen: GrapheneOS's own picker on the P9. - The Street View window as a 360° button when there is no picture (user, 2026-10-18): seen on a 4a on a saved point, the figure over "360°", and a tap opening the viewer on that block.
- The scale bar riding over the free-drive speed box instead of hiding (#848): seen on a 4a at rest, the bar through a restart and three pans. Not seen: a phone moving without a route, with Show speed on (the bar and credit raised) or off (both in place).
- A 700 ms lead-in on Bluetooth outputs (#802, 300 ms still clipped some lines for one reporter): not seen or heard; there is no Bluetooth output on the test rig. Which output the reporter uses is unconfirmed.
- A Google Takeout list import saved as it goes (#761: a 600-row list was lost when Android stopped the app part way): seen on a 4a, a 63-row test list killed at row 36 kept the 12 places found before its first save. Not seen: a real Takeout file imported again, which should take the rows it already matched with no lookup.
- A roundabout's exit step says nothing after its entry (a real drive through two roundabouts had each said twice and two lines cut off inside the ring): an engine test, which repeats that drive's lines with the rule off, and the replay audit of the drive. Not heard on a drive.
- One turn split in two by a mapped turn lane is said as one turn (a real drive heard two slight rights for one right at a divided corner): a unit test, the drive's own route (folded), and the 39 stored lines (three other folds, each a turn lane on the map). Not heard on a drive.
- The heading term and a reroute's pinned heading need 5 m/s and a 30 m fix when driving (a real drive downtown rerouted three times from GPS drift at a light): a unit test, and a replay of that drive's fixes on its first route reroutes at 52 s before and never after. Not driven.
- A recheck starts only from a real fix of 30 m or better with no off-route hits, and lands only when that fix is on the result (a recheck from a dead-reckoned point swapped in a route 185 m behind the car): a unit test of the two rules. Not driven.
- A recheck's heal no longer starts the reroute cooldown (it held a needed reroute 8 s): code only. Not driven.
- One APK per chip type, on since 2026-10-22 (
ABI_SPLITS, #680): not seen yet, a release with the five files updated through Obtainium, the F-Droid repo and the in-app updater. - Saved drives on from the first screens, deleted after 30 days (
trip_auto_clear): seen on a 4a with a fresh dev copy, the card with both switches on and both set after Continue. Not seen: a drive actually deleted at 30 days (unit test of the rule only). - The destination's side when the router gives none (a drive got "You have arrived" instead of the side), and the approach prompt saying that side instead of "ahead" (a drive heard "will be ahead", then "is on the right"): unit tests. Not heard on a drive. The 17 other languages' approach lines are built from their own stop lines and have not been checked by a speaker.
- Transit guidance in a pocket (#880): seen on a 4a with a simulated position, the card for a walk, a 42A ride with its stops left, the arrival card, End from the shade and Done. Not seen: a real ride with the screen off, the heads-up at the stop before yours (unit tests only).
- A reroute's naming snap starts the way the car points and skips vias on Google's own start (an unpinned one went round a block): a unit test of the vias. Not driven.
- A reroute whose first turn is too close to make at the car's speed is asked again once the car is past it (a start against the route got a side-street U-turn 1 m ahead at 10 m/s): unit tests, and the rule over every route swap in four recent traces fires on that one only. Not driven.
- The open router's reroute taken because Google's routes started against the car keeps Google's clock and traffic (the ETA read 9 minutes until a heal 35 s later made it 12): code only, through the helper trips with stops use. Not driven.
- Place names More and Most (#852,
MapPoiPrefs.namesSooner): not seen on a phone. Check names arriving a zoom or two sooner on Vela data, Google and Both, and frame rate with Most in a dense city on a release build (scripts/map-fps.sh).
Open from the October 2026 audit¶
Found reading everything since 0.4.1912 and not fixed yet.
- The arrow's swap to the map's own symbol during a gesture (
puckGestureSwap) is off: the car vanishes for several frames at each end of a pinch. The overlay is used throughout. - A replay seek feeds every earlier fix through the engine in one main-thread block. The catch-up shares the fix filters and the session with the main thread, so it cannot simply move to another thread, and run in slices the map and the turn card would play every earlier route and step. It needs a pass over a copy of that state, installed in one step.
- A star put on a picked suggestion or a list row, with the sheet closed before its details
load, is still kept as a point: the mark comes off only when that suggestion or list row is
opened again and its details load, not when the place is opened from Saved. A star put on a
tapped map label un-fills when the tap resolves, because the sheet takes Google's id
(
isSavedcompares ids). - Map: where the basemap itself draws the house numbers (no address overlay mounted), their box still changes its filter about every three quarters of a screen of panning at house-number zoom, and each change re-lays out every basemap tile. A free drive with no route does not hold the box back by speed as a drive with a route does. A wider box means fewer moves and more numbers drawn per frame; which wins needs measuring on a phone.
From the first drives on 0.5¶
- Match Google's line on the phone when the region is downloaded. Street names and lanes for the stretches where Google leaves the open route come from a public matching server, and a late answer leaves tile names or bare turns. The on-phone router's data could do the match with no server. Not started.
- UK fuel prices. The price file is not part of a region download, so offline it is whatever the phone last fetched, and it stops showing after two days. If the archive it comes from ends, the fallback would be the retailers' own price feeds.
- One settings panel for everything drawn on the map. The switches for what the map draws are spread over several settings pages.
- The four Texas parts need their place pack, places, basemap, road features and cells baked before the routing manifest lists them, or their downloads pull the whole state's archives.
On the radar¶
- Both-mode twins across scripts. With place icons set to Both, the twin pass compares
Google's English names with the archive's local names. Over Tokyo on an English phone 38%
link (65% under
hl=ja), so open icons draw beside their Google twin. Options: run the ambient fan-out in the region's language, or bake a romanized name into the archive. - Departure-time ETAs. Parked until someone captures one directions request from the Android Google Maps app with "Depart at" set. The history file lists what was tried.
- Ideas from the October interface round, none built: a thumbnail or a review line on result cards; the car icon changed from the drive bar; a dotted walking line from a building's door to the start of the route; a place photo from the business's own website when Google is off; Material 3 Expressive; a search bar at the bottom of the screen (#747: the hard part is a layout for the buttons and the bar that still works in landscape and with the sheets).
- iOS. Not started.
:coreis plain Kotlin and would move to Kotlin Multiplatform. MapLibre and sherpa-onnx have iOS builds. The interface would be rewritten. - A download of just the route's corridor (#732): map tiles, places and routing data within a few kilometers of a planned route instead of a whole region. Not started. The region packs are baked per region, so a corridor would be cut from them on the phone or baked on request.
- Live vehicle positions on the map (#757): Transitous serves GTFS-Realtime positions where an agency publishes them, a minority; drawn with an age label where they exist and nothing elsewhere. Not started.
- Quick search buttons of the user's own (#779): a typed query saved as a chip beside the built-in ones. Not started. The order and hidden set are keyed by the chip's query, so a custom chip would fit the same store; the open questions are the icon and what the offline store expands for a word it does not know.
?cid=links (#796):maps.google.com/?cid=<id>names nothing and Google answers it with a page, not a redirect, so opening one needs the hidden page load the reviews panel already does, reading the place off the page's final URL. Not started; short share links open.- The car's layers follow the phone's switches while connected (#749): the car builds its map once per connection, so a Traffic layer switched on afterward shows only after a reconnect. Not started.
- Contrast by day: the Contrast map colors draw as Modern in light mode. A light palette with the same intent (warm streets, roads a few steps from the land) is not drawn. Not started.
- An e-ink look (#789): a map look in black lines on white with no shading beside the five under Map look, and a high-contrast app theme without grays. The nearest today is Positron with the light theme. Not started; needs someone with an e-ink phone to look at it.
- A hiking and cycling look (#808): trails dashed by type (path, footway, track, cycleway, steps), paved against unpaved, mountain-bike difficulty, bicycle and foot access, and peaks, all of which the tiles carry today; an elevation profile for a route from the terrain tiles the shading already uses. Not in the tiles: the hiking difficulty grade, cliffs, contour lines, which need a bake of their own. Not started.
- Transit routing over OpenStreetMap's route relations (#809): the bus and tram lines mappers have drawn carry their stops and order, and Organic Maps routes over them with no timetable. Google's feed misses whole cities (Niš has its coaches and none of its city lines), and Transitous has no feed there either, so a planner over the relations in the downloaded region would answer where neither does, without times. Its own engine; not started.
- Snapping the dot to the road with no route (#803): the browse and driving-mode dot is the raw fix, and a weak head unit GPS puts it a lane or a row of houses off. Vela already finds the road under the car for the speed limit; a snap onto it while moving, released at junctions and when the fix leaves the road, is the shape. Waiting on the reporter's recording and his unit's accuracy figure. Not started.
- A swinging heading at a standstill and a speed readout that jumps at a steady speed (#818, #819, one Pixel 10 Pro, phone and car): the fix handler takes a bearing over 0.5 m/s and derives one from 3 m of movement, which may be too loose for that unit's fixes. Needs a diagnostics export of such a drive before the gates move; never seen on the project's own phones or head unit.
- Street furniture on the map (#813): benches, bins, drinking water, toilets, shelters, hydrants and the like at z17 and up. The tiles carry waste baskets, drinking water, toilets, shelters, playgrounds, pitches and bicycle parking, so those are a layer with icons of our own; benches, hydrants, fountains, single trees and retaining walls wait for our own tiles. Not started.
- Fuel and chargers along a route (#834): a count and the longest gap between
amenity=fueloramenity=charging_stationplaces near the line, on the route chooser, so two routes through sparse country can be compared before one is taken. The places come from the downloaded place pack or the open places tiles, so it works offline; the gap is measured along the line. Not started.
Big bets¶
Serving our own map tiles¶
Vela already bakes the whole world's basemap: 448 PMTiles archives, about 93 GB, counted on
the basemap-tiles release on 2026-09-25. The app draws from them wherever a downloaded
region covers the view. With nothing downloaded, the online basemap is OpenFreeMap's. Two
things stop Vela streaming its own:
- Seams. Our archives are per region and OpenFreeMap is one planet, so a pan across a boundary would swap sources mid-gesture. The fix is one planet-sized archive. A GitHub release asset caps at 2 GB and the planet is about 90.
- Release hosting is not a CDN. A map session pulls hundreds of tiles.
The missing piece is hosting: PMTiles on object storage behind a CDN, with a bill and
somebody carrying it. Do this when the project is ready to run infrastructure. With it, the
Microsoft building merge reaches streaming users and the render-time coverage gate
(runOvlGate) can be deleted, downloaded and streamed maps share one schema, and Vela stops
depending on OpenFreeMap's donated bandwidth.
A faster shortest-path search, kept as research¶
catalinb13 ported the 2025 sorting-barrier single-source shortest-path algorithm (Duan, Mao,
Mao, Shu and Yin) to Kotlin and ran it over the obf region files as a second engine (#695,
2026-10). It is kept on the branch research/sorting-barrier-sssp, unmerged, with the author's
measurements in the pull request. Measured here on the Delaware file: a route took 1.0 to 2.8 s
against 67 to 396 ms from the shipped engine, the cached graph held 404 MB of a 512 MB heap
against about 12 MB, and the graph it builds skips turn restrictions and access tags and keeps
rail and ferry lines, so a car was routed along a railway. To become an engine it would need
the search run under the lock, OsmAnd's routing rules applied when the graph is built, and a
memory gate on the file size, and it would still be slower. The interesting part is the
algorithm, and the branch is where to read it.
Richer roads¶
Google's street detail without the 3D. Roads at their real width and street names on every block shipped. What is left:
- Road markings. A developer test draws center lines, lane lines, bike lanes, crosswalks,
stop lines, turn arrows and medians from OSM tags (
core/data/PaintedRoads, dialpaintedRoads, a California bake byscripts/bake-painted-roads.sh). Open before shipping: arrows sized by eye, lane counts where OSM has none, painted islands and gores, stop lines at give-way signs, and a bake for every region. - Lanes and medians in the basemap. The OpenMapTiles schema has no lane count, so width by
lanes needs a basemap baked with an extended schema (a planetiler profile with
lanesanddivider). Downloaded regions could get it first. Online needs our own tiles.
Contributing back to OpenStreetMap¶
Vela takes its basemap, routing data, addresses, road features and half its places from OSM and gives nothing back. The API is the easy part: notes are a plain POST with no account, and editing is OAuth 2.0 with PKCE against the 0.6 API.
The firewall comes first. OSM forbids data derived from Google, and Vela shows Google's places beside OSM's. No OSM edit may be pre-filled, suggested or autocompleted from anything that came from Google, and the code has to enforce it:
- The edit path reads only fields that came from the OSM tile, the open places bake, or what the user typed.
- A place that carries a Google feature id can open a note and never a tag edit.
- The two paths share no model object.
Nothing starts until the Data Working Group has been written to. They would act if a Google-derived edit got in, and the firewall has to be agreed before there is code.
The shape follows StreetComplete: bounded questions about something the user is standing in front of, never a free-form tag editor. In order of safety: a note anywhere; "is this still here" on a place from the bake; hours, phone and website on a place with no Google listing open; a missing house number. Nothing that moves geometry. A later shape is a one-tap fix ("this place is gone") that a Vela-side service verifies against the business's own website and files under OSM's automated-edit guidelines.
Every changeset carries created_by=Vela <version> and an editable comment, uses the app's
own OAuth client, and can be undone. Testing runs against master.apis.dev.openstreetmap.org.
Notes alone could ship first.
A Google Play listing¶
Not the route for Android Auto for now. It would also reach people who never install an APK by hand.
The shape that works is a compile-time play flavor with the Google extractor left out of
the APK. What remains is a complete OpenStreetMap app: routing, open places, offline packs,
the geocoder, speed limits, the basemap, transit, road features and cameras. It loses place
pages and traffic, and the listing describes that app.
Enabling the Google half after review breaks Play policy, which is enforced on the developer account. So there are two distributions: the full app on GitHub, Obtainium and F-Droid, and the Play build with a link to the project site. An in-app APK downloader is not allowed.
Work, roughly in order:
- A flavor with the Google extractor, the hidden pages and the place-page surfaces compiled out, and search and place paths falling back to the offline stack.
REQUEST_INSTALL_PACKAGESand the in-app updater removed from that flavor.- The Data Safety form, a privacy policy URL, a content rating, and the background location declaration with its demo video.
- Package id. Play App Signing re-signs, so a Play install and a GitHub install cannot replace each other. Choose one id or two before the first upload. It cannot change later.
- Listing copy and screenshots that never imply a Google affiliation.
Opt-in telemetry¶
Diagnostics and trip recording are local, opt-in and have no upload. What is left is Vela's own traffic data: speed and route traces from opted-in users. This departs from "no telemetry, no backend", so:
- Opt-in only, with a clear consent screen, an easy off and "delete my data".
- No account. A pseudonymous device token at most. Trim the first and last 100 m or so of each trace and send speed and heading along road segments.
- It needs the first Vela backend or a privacy-preserving collector, self-hostable.
PRIVACY.mdchanges in the same commit.
Vela's own record of what is still open¶
The long game for place data, starting in the USA. Overture adds places and almost never removes one. The bake removes what Foursquare, OpenStreetMap and Wikidata mark closed (SPEC 5.2), but the Foursquare copy readable without an account stops at February 2025.
What is missing is one dated fact per place: a public record showed it open, or closed, on a date. So the project is a ledger keyed by the ids the bake already carries, fed by public registers, and read by the bake as confirm or close. Its first version adds no places.
US permit and commerce data, surveyed 2026-10-06. There is no national collection of business licenses. About 100 of roughly 19,500 municipalities publish a live feed, covering about a tenth of the population. In the order worth building:
- National sector registers, all anonymous bulk downloads: FDIC bank branches (78,061 offices with coordinates and dated closings), USDA SNAP retailers (250,628 stores with coordinates and authorization dates), CMS hospitals, NCES public schools, NPPES health providers, the IRS nonprofit file. Full US coverage, a few percent of all places each.
- State alcohol and food licenses. California and five other states publish daily lists with status and dates, and one more a weekly restaurant and hotel list. Over 40% of the population.
- Health inspections, about a fifth of the population, with trade names and coordinates.
- Big-city license feeds (New York City, Los Angeles, Chicago, San Francisco, the District of Columbia and about twenty more), one adapter each.
Licenses can confirm or close a place and cannot supply places: of the District of Columbia's 76,331 active licenses, 45% are housing rentals and 24% carry a trade name. A record is matched by address plus a loose name to a place the bake already has. An active match counts as seen alive. A removal needs an explicit closed status with a recent date and a second signal, since an expired license is usually late paperwork.
City layers often state no license, so derive closure facts from them and never republish the rows.
Open questions: whether Overture's BrightQuery rows carry a status or a last-seen date; whether to take Foursquare's current releases, which need an account and carry a logo clause; and how well one city's license feed matches the bake.
First step: FDIC and SNAP as two adapters in the places bake, confirm and close only, measured on the District of Columbia and Sacramento. National registers outside the USA (France's Sirene, the UK Food Standards Agency's list) come after the US shape is proven.
Vela traffic layer¶
Depends on the telemetry above. Opted-in traces become per-segment speed against free flow, then a traffic overlay and ETAs that need no Google. It starts as a supplement to Google's traffic and replaces it where coverage is good.
Architecture work¶
- Finish carving the large files.
NavCameraout ofVelaMapView: the follow ticker, the puck overlay and the padding and zoom eases as one class with oneframe()entry point.SearchController: query, suggestions, results, the three pickers and their gates as one tested state machine. The camera piece has to be judged on a drive. - Rules in prose become rules in code. A SPEC paragraph that describes a trap gets a unit test, a lint rule, or a type that makes the wrong state impossible, when someone touches it.
- A router with an owner. Turn-by-turn depends on the FOSSGIS community servers with no
agreement, and the only fallback is the on-phone router. A self-hosted router asked first,
with FOSSGIS as the fallback, removes that single point of failure. The engine to host is
Valhalla (decided 2026-10-03): it serves car, bike and foot, the map matching the drive
route uses, and the avoid options FOSSGIS OSRM refuses. Its tiles are memory-mapped, so a
free ARM box (4 cores, 24 GB, 200 GB disk) fits North America plus Europe. It would be the
first server Vela runs, so
PRIVACY.mdchanges the day it ships and request logging stays off. A free tier can be reclaimed, and someone has to update the tiles.
If GitHub goes away¶
GitHub is the right home today: free unlimited Actions for a public repository, releases that served 74 TB of region files at no cost, the Obtainium and F-Droid audience, and a record of keeping projects that toe a line. It is also one account that can be closed. What that would break, measured 2026-10-11, worst first.
- Installed apps. Builds before
VelaLinkscarryapi.github.comandraw.githubusercontent.comas literals, and no push can reach them. Builds from 2026-10-12 on read the settings file from the website's copy when GitHub does not answer, so a signedlinkspush there moves them. The updater reads GitHub's release JSON; Forgejo's has the same fields, GitLab's does not. - The data releases: 56,470 files, 369 GB, about 18 TB served a month. No git host replaces
that (GitLab's free storage is 10 GB). An object store with free egress (Cloudflare R2, about
six dollars a month at this size) does. Every manifest embeds absolute
github.comasset URLs, so a move is a manifest rewrite or an app-side rebase onto thereleaseslink. The 93 GB basemap and 82 GB building bakes need big-disk runners to regenerate. - The channels that pin a URL: Obtainium points at the repository, F-Droid clients at the Pages address. The F-Droid key is local, so the fingerprint survives; every user would add the new address by hand. The APK is 116 MB, over Cloudflare Pages' 25 MB file limit, so a mirror of the F-Droid repository lives on R2, not Pages.
- CI: 29 workflows, hourly to weekly crons, at least 55 runner hours a month, ten secrets (all
regenerable or kept locally), the
ghCLI in 41 scripts and workflows, artifact, Pages, CodeQL and Scorecard actions. GitLab's free tier is 400 minutes a month. Forgejo Actions runs the same YAML nearly unchanged on a self-hosted runner; GitLab means rewriting it. - Fonts on Pages fail softly: the app probes the host and falls back to OpenFreeMap's Noto glyphs. The website is on velamaps.org already; the docs would redirect there.
- Issues, pull requests, discussions, every
#123in commits and docs, the Weblate project, issue triage, download stats and the dashboard. Forgejo's importer keeps issue numbers.
Not touched: the three signing keys (local only), velamaps.org, OpenFreeMap, the Piper voices (another account's releases), the donation page.
Insurance, cheapest first. The first is built (2026-10-12: the site workflow publishes
calibration.json and its signature to velamaps.org, and CalibrationStore reads that copy
after GitHub's); the rest are not:
- A second settings-file address off GitHub (velamaps.org), published by the same workflow
as the site, so a closed account becomes one signed
linkspush. Built. Not yet seen: an app reading the copy with GitHub blocked. - The F-Droid repository mirrored on R2 behind a velamaps.org address, listed in the index as a mirror so clients fail over by themselves, and in the README.
- Manifest asset URLs rebased app-side onto the
releaseslink, so a storage move is a push, not a rebake. - A weekly export of issues and pull requests to R2.
- A GitLab pull mirror of the git history.
Not planned¶
Asked for, possible, and not being built. Each says what it would take.
- An arrival time fitted to the driver's own pace (under the limit, or 5 or 10 over). Vela's times are Google's: one figure for the route and colored congestion stretches, with no speed per road, and Vela knows the limit only of the road the car is on. A driver's pace matters only where traffic is free, so a flat percentage is wrong in traffic, and the two-minute re-check pulls the time back to Google's. It becomes possible when Vela computes its own route times (Less to Google, routes computed on the phone).
- Android 7. The minimum is Android 8.0 (
minSdk26). No library is in the way: of the 85 in the build the highest minimum is Android 7.0 (androidx.webkit), Cronet needs 6.0 and sherpa-onnx 5.0. The app's own code is:java.timein 14 files (it needs the build's back-port switched on), notification channels, the foreground service start, picture in picture, pinned shortcuts, audio focus and vibration calls with no version check, and a launcher icon that exists only in the Android 8 form. After that come a lint pass at the lower minimum, an Android 7 emulator in the smoke test, and a look at how Google's pages render in a WebView that old.
Not going to happen¶
These need a Google login or a Vela server, and the project's promise is that neither exists.
- Contributing reviews, photos or map edits to Google. Needs a Google account.
- Live location sharing and shared ETAs. Needs a rendezvous server.
- Location history or a timeline. An anti-goal.
- Live "busier than usual" popular times. Google strips the live histogram from every anonymous request. The typical-week bars are all that is available.
- A shared log linking Google listings to open place ids (declined 2026-09-18). It saves one request on a first tap, and the place is fetched anyway for its rating, hours and photos. It costs a backend, and a contribution channel that shows which places a user tapped.