liseur/DEVELOPER.md

Building

What to build next, and what has been ruled out, is in docs/ROADMAP.md.

What each supported server exposes, what is implemented against it, and why the remaining gaps are or are not fixable is in docs/SERVER_CAPABILITIES.md.

This project uses the Gradle wrapper, so you don't need Gradle installed separately: just a JDK 17+ and the Android SDK (command-line tools are enough; compileSdk/targetSdk 37 requires a reasonably recent SDK Manager package list).

./gradlew assembleDebug    # unsigned debug APK, installable as-is
./gradlew assembleRelease  # minified release APK
./gradlew bundleRelease    # minified release AAB, for Google Play only

Output APKs land in app/build/outputs/apk/{debug,dev,release}/, and the bundle in app/build/outputs/bundle/release/.

The debug build is signed with the standard Android debug key, so app-debug.apk can be installed directly with adb install or by sideloading. The release build is unsigned by default so the project builds out of the box on any machine/CI; see below if you want a signed release build.

The dev build, for testing on a phone you use

The debug build carries the production applicationId, so installing it on a phone that already has Liseur on it replaces Liseur. The library that goes with it — reading positions, highlights, notes, sessions — does not come back.

The dev build type exists so that does not have to happen:

make dev             # ./gradlew assembleDev
make dev-install     # adb install -r, beside the real app
make dev-run         # install and launch
make dev-uninstall
make dev-logcat      # logcat filtered to its pid

It is the debug build with two differences, both deliberate:

• applicationId is com.chmouel.liseur.dev. Android keys an app's storage off that, so the dev build gets its own Room database, its own DataStore and its own SAF grant. It cannot read or damage the real app's library, and the first thing it will ask for is a folder, because it genuinely has none. Connecting a server is likewise a fresh connection, with its own device key.
• The launcher name is "Liseur (dev)" and the icon field is retinted, so the two are told apart before either is opened. That is the whole of app/src/dev/res/.

The APK lands in app/build/outputs/apk/dev/app-dev.apk, debug-signed like app-debug.apk.

release is not involved anywhere in this, which is the point: F-Droid rebuilds that build type byte for byte from the tag, and nothing here changes what it produces. debug is not involved either, so make install, make run, make reset, hack/screenshots, hack/e2e-* and the scenarios under tests/ keep the package name they hardcode and go on working against the emulator unchanged.

One thing to watch if you drive it by hand: adb shell am start -n com.chmouel.liseur.dev/.MainActivity does not work. The .Class shorthand is relative to the application id, and the dev build's no longer matches the namespace the class lives in, so the component has to be spelled com.chmouel.liseur.dev/com.chmouel.liseur.MainActivity. The Makefile already does.

Signing a release build (optional)

The real signing key for published releases lives in pass:

Entry	Contents
android/liseur.keystore.p12	the PKCS#12 keystore, base64
android/liseur.keystore-password	store and key password
android/liseur.keystore-alias	key alias (liseur)

To build a signed APK locally with it, write a keystore.properties (gitignored, per-developer) pointing at a decoded copy. Keep that copy in persistent user data rather than /tmp, which may be cleared between builds:

signing_dir="${XDG_DATA_HOME:-$HOME/.local/share}/liseur/signing"
install -d -m 700 "$signing_dir"
umask 077
pass show android/liseur.keystore.p12 | base64 -d > "$signing_dir/release.p12"
{
  echo "storeFile=$signing_dir/release.p12"
  echo "storePassword=$(pass show android/liseur.keystore-password)"
  echo "keyAlias=$(pass show android/liseur.keystore-alias)"
  echo "keyPassword=$(pass show android/liseur.keystore-password)"
} > keystore.properties

If keystore.properties exists but its storeFile has been removed, Gradle stops during configuration with recovery instructions. Run the commands above again to restore the decoded keystore and refresh the properties file. Delete keystore.properties only when you deliberately want the unsigned build used by clean checkouts and F-Droid.

hack/release performs this check before a release and runs the same recovery automatically when the properties file or decoded keystore is missing. It leaves an existing usable custom signing path unchanged.

Contributors without access to that key can generate their own instead. Any key produces an installable APK; it simply won't update over one signed with the release key:

keytool -genkeypair -v -keystore /path/to/your.p12 -storetype PKCS12 \
  -alias liseur -keyalg RSA -keysize 4096 -validity 10950

Either way ./gradlew assembleRelease picks the file up automatically; without it the release build is simply unsigned.

Testing

./gradlew testDebugUnitTest   # JVM unit tests
./gradlew lintDebug           # Android Lint (0 errors required)

There is no instrumented/emulator test suite. Reader interactions (gestures, immersive mode, process-death restore, rotation) are verified manually on a booted AVD.

The upload end-to-end check

Sending a local book to liseur-sync is the one path where a unit test proves almost nothing: what can go wrong is the shape of the whole round trip, not a function. hack/e2e-upload drives it through the app's own screens against a real server and then reads both sides.

hack/e2e-upload -u http://10.0.2.2:8686 -t <token> -d /srv/books

-u is the server as the device sees it, so an emulator wants 10.0.2.2 and never 127.0.0.1. -t is a device token holding at least sync, library-read and library-upload. -d is the watched folder's root on this machine, which is how the script sees what landed. The folder has to accept uploads already:

liseur-sync admin folder-uploads <folder-id> on

It asserts four things, and each one is a bug that actually happened: the library-upload scope reaches the app as can_upload, the files appear in the folder, no uploaded book had its url rewritten (that key is what every reading position hangs off), and no book came back from the following catalog pass as a second row. That third one is worth the whole script: a book uploaded from the device was being deleted by the next refresh, and its reading position with it.

The other half of the capability is refusing, and -r checks it:

hack/e2e-upload -r -u http://10.0.2.2:8686 -t <token> -d /srv/books

A server says no in two places. A token without the scope is refused on sight, and the action is never offered. A token that holds the scope but finds every folder closed can only be refused by trying, and the app has to remember the answer; otherwise it offers, once per book, an action that silently fails every time. Either way -r asserts the app ends up not offering, no book was linked, none left the shelf and nothing was written into the folder. Run it with the folder still closed, then folder-uploads <folder-id> on and run the check above, and you have covered both answers with one server.

Nothing about it is mocked. Start with a clean shelf (hack/reset-books then adb shell pm clear com.chmouel.liseur), or the counts it compares are counting an earlier run.

Checking a folder's storage permission

Deleting a book's file needs write access to the folder it lives in, and that access is granted once, by the system picker, when the folder is added. A folder added by a version of Liseur older than "deleting a book deletes the book" was only ever granted read.

What the system actually holds is worth checking directly rather than inferring from behaviour:

adb shell dumpsys activity permissions | grep -A2 targetPkg=com.chmouel.liseur

mode=0x1 is read only, mode=0x3 is read and write, and persisted says which of those survives a reboot. Adding the same folder again through the picker upgrades a read-only grant in place, which is what the "add it again to grant deletion" wording in delete_local_failed is telling the user to do.

Worth knowing: on AOSP 16 deletion succeeds even from a read-only persisted grant, so a stale grant does not reproduce the failure there. The fallback in LocalLibraryRepository.addFolder and the message that goes with it are for the devices where it does.

Checking the local network permission

Android 17 blocks an app targeting SDK 37 from reaching addresses on the phone's own network until the reader allows it, and it blocks below the HTTP client: a LAN server does not refuse a connection, it swallows it until the timeout. That is #195, and ADR-0028 explains the shape of the answer.

What the system holds is worth reading directly:

adb shell dumpsys package com.chmouel.liseur | grep -i local_network

An image older than API 37 can be made to behave like one, which is the only way to see the failure without an Android 17 device:

adb shell am compat enable RESTRICT_LOCAL_NETWORK com.chmouel.liseur.dev
adb shell am compat reset RESTRICT_LOCAL_NETWORK com.chmouel.liseur.dev

Note that granting any permission in the NEARBY_DEVICES group grants this one with it, so a phone that has been asked about nearby devices for another reason will not show the prompt.

Why the reading stats say "Counting this device only"

The dashboard combines every device that has signed in to liseur-sync, but only when the merge can be proved exact. When it cannot, the screen silently shows what this device counted and says so. That is deliberate: ADR-0021 would rather report a smaller true number than a plausible wrong one, and a reader is never shown an error about it.

The reason is written to the log instead, so the answer is one command away:

adb logcat -s liseur-sync-insights

Every refusal prints Statistics count this device alone: <reason>. The reasons are prose, not codes, and name the exact thing that could not be established, from the account having no identity to check a reply against, through a book resolved to a different work since it was sent, to a reason the server itself gave.

The one worth recognising is the server could not place this device's evidence (candidate_payload_mismatch). It means this app rebuilt the description of a sitting and the server disagreed with its own stored copy. In practice that means the device identity behind those uploads changed, since the server folds the device id into the fingerprint it compares. A reconnect with a password keeps the stored device id and is fine; a pasted token names nobody and gets a fresh one.

Some background on why any of this needs rebuilding. Sessions have been uploaded since v0.6.0, but the table that retains the exact bytes of each request only arrived with the dashboard. Every sitting older than that is of unknown standing: the upload flag does not settle it either, because disconnecting an account clears the flags while the server keeps the sessions. So the sitting is described again from the stored row and offered as a candidate, and the server rules on it. That rebuild is byte-identical to the original because the payload is a pure function of the row, and the only field added since, active_ms, is withheld from exactly these sittings.

Two cases are settled without asking. A sitting recorded since the evidence table existed and never sent is counted here, because a request is always written down before it goes out. A sitting of a book the server has no confident name for is also counted here, because sending reading up requires such a name and the upload query joins on it, so that sitting has never once been selected.

Reproducing the upgraded state on an emulator is the quickest way to exercise all of this: flag the sittings, drop their retained requests and clear their upload flags.

adb shell "run-as com.chmouel.liseur sqlite3 databases/liseur.db \
  'UPDATE reading_sessions SET legacy_evidence_unknown = 1, uploaded_at = NULL; \
   DELETE FROM session_transmission;'"

The screen should still read "Counting all your devices", and the sitting count should equal this device's own, not that plus the server's. Never do this on a phone somebody reads on.

Releasing

versionName, write the F-Droid changelog, run the tests, lint and a release build, commit, tag and push, publish the GitHub release, and update the F-Droid submission.

Run it from a clean, up-to-date main branch, with nothing after it:

hack/release

It shows what has landed since the last release, grouped by commit type, says so when the screens have changed since the screenshots were last taken, offers the next patch, minor and major version, and opens an editor on a changelog drafted from those same commits. Correct the draft, save, and confirm.

The version can also be given outright, which is what CI and scripts want:

hack/release -n 0.2.1 "Fix page fitting on tall screens."
hack/release --fdroid-only 0.2.1     # re-run just the F-Droid step
hack/release --no-play 0.2.1 "..."   # leave Google Play out of this one

-n (--non-interactive, --yes) answers the confirmation prompts, and any run with nothing on stdin needs it: the prompt reads end-of-file as no and the release stops without having done anything. That applies to --rc too.

It refuses to run on a dirty tree, off main, out of sync with the remote, on a version that is not newer, or with release notes over the 500 characters F-Droid allows. An interrupted run can be resumed by invoking it again with the same version.

Pushing the tag is what starts .github/workflows/release.yml, which builds and signs the APK in the release GitHub environment and attaches it to the release. Signing is what that environment must hold: LISEUR_KEYSTORE_BASE64, LISEUR_KEYSTORE_PASSWORD, LISEUR_KEY_ALIAS, LISEUR_KEY_PASSWORD and, for the release notes, GEMINI_API_KEY. Without them there is no release.

LISEUR_PLAY_SERVICE_ACCOUNT_JSON is the odd one out. It buys the third channel rather than the release itself, and the workflow skips Google Play entirely when it is absent. hack/release uploads it with the rest all the same, so in practice a release goes to Play unless you say otherwise:

hack/release --no-play 0.9.0 "..."

The script records that per-release choice in the annotated tag, so an already configured Play credential cannot override it. Deleting the secret by hand is neither necessary nor useful: a later ordinary release will restore a missing credential from pass.

hack/release creates the environment and uploads whichever secrets are missing straight from pass, so an unlocked password store is the only setup a fresh clone needs. To refresh them all, after rotating the key for instance:

hack/release --sync-secrets

Signing a build on your own machine is a separate matter, covered under Signing a release build above.

Release candidates

A release candidate is the same build, signed with the same release key, put where a handful of people can install it before the final release:

hack/release --rc             # the next patch, numbered -rc.1, then .2
hack/release --rc 0.16.0      # a version you are working towards
hack/release --no-play --rc   # skip Google Play testing tracks

It publishes a GitHub prerelease with the signed APK attached, uploads the app bundle to Google Play's internal and closed (Testing) tracks, and stops there: no F-Droid merge request, no changelog to write, and no Google Play production upload. Before the tag is pushed, the generated RC commit must pass the JVM tests, Android lint, and the release build in its detached worktree. The reproducibility check remains part of the final release path. Pass --no-play to leave Google Play out.

The signature is what makes it worth doing. It is the one F-Droid publishes under, through the dual-signing flow, so the APK installs straight over a copy that came from F-Droid, with no uninstall and no lost library, and F-Droid offers the next real release over it afterwards, as an ordinary update.

That last part is only true because the RC takes a versionCode and the next real release lands above it. hack/release counts the next code from the highest one across main and every tag, so an RC at 32 pushes the following real release to 33, and F-Droid sees an upgrade. Nobody who installed an RC is stuck, but they are on it until the next release goes out, since F-Droid will not offer a lower versionCode.

Nothing lands on main. The commit that bumps the version is reachable only through its tag, so main stays a history of real releases and the next one still bumps from the last real version. Run it from any branch, as long as the tree is clean.

F-Droid is told nothing about an RC, but it reads the public tags. v0.9.4-test.1, the old prerelease spelling, was once picked up by its checkupdates bot, which opened a merge request proposing it as the current version. The fdroiddata metadata therefore filters what the bot looks at, with UpdateCheckMode: Tags ^v[0-9.]+$; neither -rc.N nor the legacy -test.N spelling matches. Its AutoUpdateMode: None also means builds are submitted only by hack/release, but the tag filter remains necessary: the update checker and build submission are separate mechanisms. Before any final or RC tag is pushed or submitted, hack/release runs hack/verify-fdroid-tags against that live metadata. It fails closed if a final tag would be missed or either prerelease spelling would be discovered, so a future metadata or tag-format change cannot silently put an RC on F-Droid.

Run make verify-fdroid-tags for the deterministic local regression checks, or run hack/verify-fdroid-tags v0.15.0 v0.15.0-rc.1 v0.15.0-test.1 to check the live fdroiddata filter directly. A change to the RC tag format or the UpdateCheckMode pattern must update and rerun this policy check.

To install one, download the APK from the release page and adb install -r, or hand it to whoever is testing.

GitHub release pages are retained more narrowly than their tags. After an RC is published, hack/prune-prereleases deletes every legacy -test.N release page and keeps only the highest -rc.N page for each target version. Its APK goes with the page. The tags are never deleted: they keep versionCode allocation monotonic, number later candidates correctly, and remain burned names under GitHub's immutable-release rules.

Google Play

Play is the third channel, after the GitHub release and F-Droid, and it is deliberately the least load-bearing of the three. The same tag that publishes the release also builds an app bundle and pushes it to the internal testing track, then promotes that same build to the closed track named Testing:

make bundle     # ./gradlew bundleRelease, for Play only

The second track is not decoration. Internal testing holds a hundred hand-listed addresses and hands out its own opt-in link; the closed track is the one testers are recruited into (issue #62), the one https://play.google.com/apps/testing/com.chmouel.liseur leads to, and the only one whose opted-in testers count towards the twelve Play wants for fourteen days before a personal developer account may publish. Uploading to internal alone, which is what happened up to 0.10.0, leaves everyone who followed the instructions on whichever build the closed track was last given by hand (0.9.3, as it turned out, four releases back). hack/store-status now says so in a line when the closed track falls behind internal, because nothing else did.

Nothing about the APK path changes. F-Droid's recipe builds assembleRelease and never sees fastlane/Fastfile, no Gradle publishing plugin is applied, and the bundle is an extra output rather than a replacement, which is also why the Play step in .github/workflows/release.yml is continue-on-error and skips itself entirely when the service account secret is absent. A fork, or a rejected upload, must not be what makes a release fail. hack/release --no-play turns that skip into something you can ask for: its annotated tag carries release:no-play, and the workflow checks that marker before it considers the stored credential.

The upload runs fastlane android internal with SUPPLY_JSON_KEY pointing at a service account credential written to $RUNNER_TEMP and removed by a trap in the same step. The credential is android/google.play.service.serviceaccount in pass, shared with the other apps on the account; it needs Release to testing tracks on Liseur under Users and permissions.

A build that is already on internal, a release whose Play step was skipped, or one that predates the promotion, can be pushed across without rebuilding:

fastlane android promote version_code:19

The version code defaults to whatever app/build.gradle.kts currently says, so the argument is only needed for an older build.

Who can actually install it. Everything above puts a build on a track; none of it lets anyone in. The tester list is console state, and these are the things that quietly turn a correct opt-in link into "this app isn't available":

• The closed Testing track needs an email list attached, holding the Google account each tester uses on their phone. hack/store-status says which of the two kinds of list the track is on, but it cannot say who is on an email list: edits.testers models only Google Groups, and its own reference says email lists are not supported by the resource. That address goes on by hand in the console, and nothing here can do it for you.
• The console accepts either an email list or a Google Group, not both. The console offers Email and Google Groups as one choice, not two, and the API has only a googleGroups[] field. Attaching a group cuts off every tester already on the email list until each of them joins it, which restarts the fourteen-day clock this whole exercise is running down. Two things are worth knowing before reaching for one anyway: the trap that made the group fail before is avoidable, because a consumer group's owner can directly add members rather than wait on join requests, but a group still buys no automation. Consumer @googlegroups.com membership has no API of any kind, Admin SDK and Cloud Identity both being Workspace-only, so it moves the clicking from one web UI to another.
• Country availability is not the reason, and has not been for as long as anyone has been asking. The track is open to 176 named countries and to the rest of the world, so there is no country left over to be missing; hack/store-status prints this outright so the question stops being guessed at. Russia is on the list and works: Google paused Play billing there in 2022 and stopped paying developers with Russian bank accounts in 2024, and both notices say in terms that free apps still download. Only North Korea and Syria are absent, which is Google's sanction policy rather than a switch in the console. This is track-level console state, and the release carries no countryTargeting of its own, which is why a supply upload has never disturbed it.
• The track's release must be completed, not draft. hack/store-status prints the status of each one.
• Adding an address is not instant. Play takes a few hours to propagate, and up to a day or two to move the opted-in count, so a tester told to try again should be told to wait.
• If the store still says the app is not available after all that, the usual cause is on the tester's side: the account signed into the Play Store on the phone is not the address that was added. A browser signed into a second Google account looks exactly the same.

The service account holds Release to testing tracks, which covers everything in this repository: uploads, promotions, and every read hack/store-status makes. It deliberately does not hold Manage testing tracks and edit tester lists (CAN_MANAGE_TRACK_USERS_GLOBAL), the permission edits.testers.update would need. Nothing automated has any business rewriting who may install the app, and while the track runs off an email list that call could not help anyway.

Only the changelog is pushed from the repository. The store listing is edited in the console, because Play holds declarations that no file here describes (data safety, content rating, target audience, app access, ads), and those have to be revisited whenever the app gains a permission, talks to something new, or changes what it stores. The privacy policy Play links to is docs/PRIVACY.md, served by GitHub Pages from main /docs; it is a published legal document, so change it in a commit and not in the console.

Promotion out of testing and in front of the public is a manual action in the console, on purpose. Nothing in this repository can put a build on the production track.

The signing key. Play App Signing holds the same key as the GitHub release, enrolled from pass through Google's PEPK tool rather than generated by Google. That means a build installed from Play and a build downloaded from the release page carry the same signature, so a reader can move between them without uninstalling and losing their library. It also means the key cannot be swapped without a rotation request. F-Droid publishes two APKs per version, one with its own signature and one with this same key grafted onto its reproduced build (see F-Droid readiness below), so a reader can cross between all three channels.

Recreating the PEPK export, should it ever be needed:

d=$(mktemp -d) && umask 077
pass show android/liseur.keystore.p12 | base64 -d > "$d/release.p12"
java -jar pepk.jar --keystore="$d/release.p12" \
  --alias="$(pass show android/liseur.keystore-alias)" \
  --output="$d/output.zip" --include-cert \
  --rsa-aes-encryption --encryption-key-path=encryption_public_key.pem

pepk.jar and encryption_public_key.pem are downloaded from the console, and the temporary directory goes away afterwards.

The two edge-to-edge advisories

Every release raises the same pair of warnings on the Play Console, and neither is a defect report:

Edge-to-edge may not display for all users.

Your app uses deprecated APIs or parameters for edge-to-edge.

The first is attached to every app targeting SDK 35 or later. It asks us to test, not to change anything. The app has drawn edge-to-edge since it started targeting 35: MainActivity and ReaderActivity both call the no-argument enableEdgeToEdge(), no theme sets android:statusBarColor or android:navigationBarColor, and the manifest carries no opt-out.

The second is static analysis of the bundle's bytecode, and it does not say whose bytecode. Ours calls none of the deprecated setters. Every flagged reference belongs to androidx:

Artifact	setStatusBarColor	setNavigationBarColor	setDecorFitsSystemWindows	setSystemUiVisibility
androidx.activity:activity	4	4	4	0
androidx.core:core	1	1	4	7

They are the version-guarded backports inside androidx's own EdgeToEdge, WindowCompat and WindowInsetsControllerCompat, which are the classes Google's own migration guide sends apps to. There is nothing to migrate away from, and hand-rolling around them to quieten a console page would be strictly worse. The warning stays until androidx changes, so treat it as noise rather than as something to fix before a release.

Re-check the claim, rather than trusting this table, if androidx moves:

activity_version=$(sed -n 's/^activityCompose = "\(.*\)"/\1/p' gradle/libs.versions.toml)
unzip -p "$(find ~/.gradle/caches/modules-2/files-2.1/androidx.activity/activity/"$activity_version" \
  -name '*.aar' -print -quit)" classes.jar > /tmp/a.jar
unzip -p /tmp/a.jar '*.class' | grep -ao setStatusBarColor | wc -l

What the first advisory is good for is prompting an actual look. Doing that once turned up two real inset bugs the console never mentioned: the reader's search results were laid out 820px beneath the keyboard where nothing could reach them, and neither the search nor the contents screen kept clear of a display cutout in landscape. Both screens pinned contentWindowInsets to WindowInsets.systemBars, which omits the IME and the cutout; safeDrawing covers all three. An emulator with a cutout is the way to see it:

adb shell cmd overlay enable com.android.internal.display.cutout.emulation.tall
adb reboot   # the display picks the cutout up on boot, not on enable

Release notes

Two things are written for every release, and they are not the same thing:

• The F-Droid changelog is fastlane/metadata/android/en-US/changelogs/<versionCode>.txt. Written by hand, capped at 500 characters, and passed to hack/release as the release notes argument. This is what F-Droid shows.
• The GitHub release notes are generated during the release by hack/generate-release-notes. It asks Gemini to turn the commits since the previous tag into something a reader would want to read, using the hand-written changelog as the summary to lead with. It also gives Gemini the associated pull-request descriptions, links changes back to those PRs, and carries a relevant screenshot over from a PR body when one is available. Those screenshots are re-sized to a fixed width (SCREENSHOT_WIDTH in the script) before the notes are written, so a full-size PR image cannot blow the page apart.

Without GEMINI_API_KEY, or if the request fails, the generated notes fall back to the hand-written changelog, so a release is never held up by this. The key comes from pass under google/gemini-api and is uploaded to the release environment by hack/release --sync-secrets. The generator keeps the prompt compact by omitting commit diffstats and retries transient Gemini or network failures with bounded backoff before using that fallback.

The notes can be rewritten after the fact: the body of a release stays editable even though its tag and assets do not:

GEMINI_API_KEY=$(pass show google/gemini-api) \
  hack/generate-release-notes --output notes.md v0.2.0
gh release edit v0.2.0 --notes-file notes.md

Never delete a final release or release tag

Releases are immutable once published, and that goes further than the assets: a tag that has carried a release can never be used again, even after deleting both the release and the tag, and even with the feature turned off. GitHub does this so that a trusted artifact can never be swapped for another under the same name.

So a final release that went out wrong is not fixed by deleting it. Bump the patch version and release again. v0.1.0 was burned exactly this way and is why the first published version is 0.1.1.

The deliberate exception is the GitHub release record for a prerelease. hack/prune-prereleases removes obsolete RC pages and their APK assets to keep the releases list useful, but never removes the tag. The name remains burned and the source commit remains reachable, which is also why neither the script nor the workflow ever uses gh release delete --cleanup-tag.

What F-Droid checks

Updating the submission is not the end of it. Pushing to the metadata merge request starts a pipeline on fdroiddata, and that pipeline can fail long after hack/release has finished and reported success.

The check that catches people out is fdroid rewritemeta. It reformats metadata/com.chmouel.liseur.yml and fails if the result differs from what is committed, byte for byte, trailing newline included. It is not a linter with opinions to argue with: the file has to be what it would have written.

Every release for a month failed this job while the builds themselves passed. hack/release had been sending the metadata through jq as --arg content "$(cat file)", and command substitution strips trailing newlines, so what arrived ended mid-line. Use jq --rawfile, which reads the file as it is. The same trap is waiting in any script that sends a file through a JSON API.

hack/release now waits for the metadata checks and stops the release naming whatever failed. What it does not wait for is fdroid build, which compiles the app from source, and the check apk that follows it: about twenty-five minutes between them. Those are left running and the pipeline is linked in the output, so look at it before assuming a release landed.

A failure in fdroid build usually means reproducibility, which hack/verify-reproducible will reproduce locally. See F-Droid readiness below.

Store assets

The screenshots and the icon are regenerated rather than maintained by hand:

hack/screenshots --setup    # build the demo shelf first, then capture
hack/screenshots            # capture from a device already set up
hack/screenshots --setup-only   # build the shelf and stop, to check it
hack/screenshots --class tablet  # file the captures as a tablet set
hack/screenshots --no-dictionary # skip the one capture that needs the network
hack/icon                   # fastlane icon.png, from the vector drawables
hack/feature-graphic        # fastlane featureGraphic.png, from the brand emblem

hack/screenshots drives a connected device through adb and writes both docs/screenshots and the matching fastlane directory. It finds controls by what they say rather than by where they sat when it was written, so a moved button is something it waits for and fails on, not a tap into empty space.

There are two sets, because fastlane and F-Droid publish phoneScreenshots and tenInchScreenshots separately and the README embeds the phone set by name. Which one a run writes is decided by how wide the device is (under 600dp phone, above it tablet), so a capture against a tablet cannot quietly overwrite the phone images. Pass --class for a device whose shape does not match how its pictures should be filed.

The phone set is the full tour, seventeen screens, gathered on docs/SCREENSHOTS.md; the README shows three of them. The tablet set is three pictures of what a phone cannot show (two columns, the control that chooses them, and a shelf with room on it) and lands in docs/screenshots/tablet. There is no point photographing the settings screen twice, and the search and dictionary steps that make the phone run slow are skipped, so a tablet run takes a few minutes.

--setup builds the shelf from nothing: it downloads a handful of Standard Ebooks public domain editions (their cover art is what makes the library screens publishable), pushes them to the device, grants the folder through the real picker, and leaves the first book part-read with three highlights, a note and a bookmark on it. Each of those is checked against the database afterwards, because a highlight that quietly failed looks exactly like one that worked. One of the three marks runs down the page rather than along a line, because the notebook only offers to open a mark out once it is longer than four lines. --setup also writes six weeks of sittings into reading_sessions with run-as sqlite3: the stats screens count from the moment a book is first opened, and a shelf seeded an hour ago has nothing on them. Run it once; later runs can drop --setup and take about ten minutes.

The definition card is the one capture that needs the device to reach something: the run turns the lookup on in Settings, presses words until one comes back with senses on it, and puts the switch back where it found it. It used to keep the previous image and carry on when nothing came back, which is how a stale picture survived several releases. It now fails. --no-dictionary is how to say an offline run was expected.

--empty is its own mode, and short: the empty library is the one screen the tour cannot reach, because the tour needs a shelf with books on it. It wipes app storage, photographs what a new reader sees, and stops. Run it whenever the empty state changes; the rest of the tour would only put the demo shelf back.

Everything that gets published is in the light theme. A dark screenshot in a store listing reads as the app looking like that, rather than as the app being able to; the dark theme earns more as a line in the description than as one picture in six that matches none of the others. The script still captures 11-reading-dark and 17-empty-library-dark, which are useful to look at, but it just does not file them with fastlane or the README.

Sign out of calibre-web first unless your server holds only books you would publish a picture of. The script itself never writes to one.

Look at the images before committing. A screen can be found and still be showing the wrong thing:

montage docs/screenshots/*.png -tile 6x2 -geometry 320x+6+6 /tmp/sheet.png

Architecture

See AGENTS.md for the layered package layout and project conventions. Key decisions:

• Readium Kotlin Toolkit (readium-shared, readium-streamer, readium-navigator, readium-opds) does EPUB parsing, rendering, and OPDS feed parsing. readium-lcp is deliberately excluded, since it depends on the proprietary liblcp, incompatible with F-Droid.
• calibre-web integration is two protocols: OPDS for browse/search/ download, and the Kobo sync protocol (/kobo/<token>/v1/...) for reading-position sync, exchanging percentage progression like KOReader does.
• Komga integration is Komga's own REST API: POST /api/v1/books/list to browse, GET /api/v1/books/{id}/file to download, and GET/PUT /api/v1/books/{id}/progression to sync a full Readium locator rather than a percentage.
• One server is connected at a time. data/remote/ holds provider-neutral contracts (CatalogSource, FileSource, ServerSetup, PositionSync); data/calibre/ and data/komga/ implement them, and RemoteRouter picks the implementation from the connected server's ServerKind. domain/ReadingStateMerge.kt is shared by both, so the conflict rules are written once.
• Single :app module, manual DI composition root, ViewModel + StateFlow, Room + DataStore for persistence.

calibre-web protocols

Verified against a real calibre-web install (behind Caddy + Cloudflare) and against the calibre-web source (cps/opds.py, cps/kobo.py, cps/kobo_auth.py, cps/services/SyncToken.py).

OPDS catalog

• OPDS 1.2, Atom XML, served under /opds (application/atom+xml;profile=opds-catalog).

• Auth is HTTP Basic on every route; a 401 carries WWW-Authenticate: Basic realm="Authentication Required". Anonymous browsing is a server-side option.

• Navigation feeds: /opds/new, /opds/discover, /opds/rated, /opds/hot, /opds/author, /opds/publisher, /opds/category, /opds/series, /opds/ratings, /opds/formats, /opds/language, /opds/shelfindex, /opds/readbooks, /opds/unreadbooks.

• /opds/books is a letter index; the full list is /opds/books/letter/00.

• Paging is ?offset=N, 60 entries per page, with rel="next" and rel="prev" links. Search results are not paged.

• Search is /opds/search/{terms} or /opds/search?query={terms}; the OpenSearch description is at /opds/osd.

• An acquisition entry carries the calibre UUID as <id>urn:uuid:...</id>, covers as /opds/cover/<int id> (integer, not UUID) and a download link:

  <link rel="http://opds-spec.org/acquisition"
        href="/opds/download/74/epub/" length="156172" title="EPUB"
        mtime="2026-07-26T10:26:49+00:00" type="application/epub+zip"/>
  

• Downloads honour Range (206) and send ETag, Last-Modified and a UTF-8 Content-Disposition filename, so resumable downloads work.

• A user without the "Allow Downloads" permission gets 401 on /opds/download/... (403 on the web UI route) while browsing keeps working. The app must recognise that case and tell the user to enable that permission for their account in calibre-web, rather than showing a generic failure.

Kobo sync

Off by default; the admin enables it in Feature Configuration, then each user creates a token from their profile page ("Kobo Sync Token"), which yields a base URL of the shape https://host/kobo/<32 hex chars>. The token is the only credential; it never expires and there is no user agent or device check.

• GET /v1/library/sync returns a JSON array of entities: NewEntitlement, ChangedEntitlement, ChangedReadingState, DeletedTag, ... Each entitlement bundles BookEntitlement, BookMetadata and ReadingState.

• Paging is via the x-kobo-synctoken header (base64 JSON of per-table timestamps); the response repeats it, and sends x-kobo-sync: continue when more pages remain. Sending the previous token returns only what changed since.

• GET|PUT /v1/library/<uuid>/state reads and writes the reading position. The PUT handler indexes its keys directly, so CurrentBookmark, Statistics and StatusInfo must all be present (any of them may be null); omitting one returns 400:

  {"ReadingStates": [{
    "CurrentBookmark": {"ProgressPercent": 42,
                        "ContentSourceProgressPercent": 42,
                        "Location": null},
    "Statistics": null,
    "StatusInfo": {"Status": "Reading"}}]}
  

  Status is ReadyToRead, Reading or Finished. Writes are last-write-wins, with no conflict detection, so the client compares LastModified itself. A null Location leaves the stored location untouched rather than clearing it.

• Position mapping: ProgressPercent / 100 maps to a Readium Locator.locations.totalProgression. Location.Value is a kepub span id (kobo.7.1) that has no Readium equivalent, so it is read but not written.

• Books are offered as KEPUB when the server has kepubify configured (DownloadUrls then lists KEPUB only). KEPUB is EPUB3 with extra spans, and Readium opens it fine.

• Behind a reverse proxy, download URLs are built from the forwarded host; a Cloudflare/Caddy install can still emit http:// URLs, so the client must rewrite the scheme to match the configured base URL.

• Deletions in calibre are not propagated (only archived books are, as IsRemoved), and users can restrict syncing to selected shelves, which makes an empty sync legitimate.

Komga protocol

Verified against a real Komga install (API 1.25.0) with an API key, and against BookLifecycle.kt on master. Every rejection below was provoked deliberately rather than read off the schema.

• Auth is X-API-Key: <key>, created in Komga's web UI. Identity is GET /api/v2/users/me -> {id, email, roles[]}; there is no /api/v1/users/me. FILE_DOWNLOAD in roles means the account may download. Users paste the API-key page address, so setup reduces a pasted URL to its origin.
• Browse is POST /api/v1/books/list?page=&size=&sort= with an EPUB + READY condition; GET /api/v1/books is deprecated since 1.19.0. Each entry carries readProgress {page, completed, readDate} inline, which is the change detector: a routine sync costs one request.
• Search is the same endpoint with fullTextSearch set, a sibling of condition.
• PUT /progression validates, in order: modified strictly after the stored readDate (else 409); href, fragment stripped and URL-decoded, exactly matching an internal EPUB file name, with no leading slash (else 400); and locations.progression, which is required. Our totalProgression is ignored and recomputed.
• There is no page-based fallback. PATCH read-progress {"page": N} is rejected for reflowable EPUB ("not Divina compatible"); only {"completed": true} works. On a 400 the client instead fetches GET /api/v1/books/{id}/positions and snaps to the nearest position Komga already knows, keeping the position rather than coarsening it. That index is ~330 KB for an ordinary book, so it is only ever fetched after a rejection.
• 409 is not a failure. The server holds something at least as new; the row stays dirty and the next run pulls and reconciles.
• GET /progression answers 204 with an empty body when there is no progress. A 404 means the book is unknown, and is a failure.
• modified comes back with a local UTC offset and is double-offset, so it must not be used for ordering; readProgress.readDate round-trips exactly and is what the sync orders by.
• Deleting a book from the server is admin-only (DELETE /books/{id}/file), so that action is hidden for Komga.

liseur-sync protocol

A second kind of partner, and a different shape from the two above. calibre-web and Komga each hold one current position per book and answer "where am I"; liseur-sync holds an append-only log and answers "what has happened since seq". It holds no books at all, which is what lets it sync a book that came off an SD card.

• Auth is Authorization: Bearer <token>. POST /v1/login with a username and password mints device tokens through POST /v1/tokens; one scope per token, so signing in asks for two, sync and read-insights, and a reader who pastes a token made elsewhere usually has only the first. Statistics are simply absent then.
• The cursor is the only irreplaceable state. Everything else can be asked for again, but ops behind sync_account.cursor_seq cannot. So a page from GET /v1/changes?since= is written and the cursor advanced in one transaction, never the other way round. A cursor that has fallen below the server's compaction horizon gets 410 resync_required; the answer is GET /v1/heads and its snapshot_seq.
• Ids are derived, not drawn. The server treats op_id and session_id as idempotency keys and compares the whole payload behind each: same id and same payload is duplicate, same id and a different payload is a conflict. So the id is UUIDv3(deviceKey|workId|revision) and every payload field comes from stored state: client_ts is reading_progress.updated_at, never the clock. A push interrupted by a dead network is simply repeated. The server does not check that the id is a UUIDv7; it is opaque, up to 64 characters.
• POST /v1/works/resolve takes every identifier at once (sha256:, pmd5: (KOReader's partial MD5), dc: and ta:) and registers all of them against whichever matched, which is how a re-encoded copy and the original converge. 409 means they named two different works; the server changes nothing and merging is left to the reader.
• ta: normalisation is an interoperability contract and is not in the schema. The server matches ta aliases by exact string and computes nothing itself. WorkIdentifiers.titleAuthor defines it as fold(title)|fold(author), where folding is NFKD, strip \p{Mn}, lowercase, non-alphanumerics collapsed to single spaces, trimmed. Any other client must agree exactly or it will silently fail to match.
• A locator over 16 KB is dropped and the progression sent on its own. Failing the push outright would leave the other device with no idea where the reader is, which is far worse than reopening at a percentage.
• A book resolved today has all of its history behind the cursor, so a newly named book is seeded once from GET /v1/works/{id}/positions?limit=1.
• POST /v1/sessions takes closed sessions only, as progression fractions. Never send page numbers: a page is a property of one rendering of one edition at one type size, and the server derives pages itself when it knows the edition. idle_ms is always zero here and honestly so: time is counted only while the reader is in the foreground, so time spent elsewhere is already absent rather than included and subtracted.
• Statistics (/v1/insights/*) are decoration. Every failure is null and silent, and a null eta_seconds is carried through untouched: no estimate beats an invented one.
• Uploading is opt-in twice over. The server advertises the library-upload scope on GET /v1/token and marks the folders that take uploads in GET /v1/folders; without both, the action is not offered at all, which is how an older server needs no version check. POST /v1/folders/{folder}/books is multipart/form-data, keyed by the file's SHA-256, so a repeated upload answers 200 duplicate and stores nothing twice. 202 means the bytes are safe but the server had not catalogued them yet; the worker simply asks again, and the digest makes the second ask free.
• What follows an upload is adoption, not replacement. The local row keeps its own url and gains remote_uuid and download_href (BookDao.linkToRemote). Rewriting the URL to the server's spelling would take every reading position, annotation and session with it. Because the catalog reads what the library holds once, before its walk begins, both sides guard against the book being introduced twice: the catalog re-asks about the ids on a page it has not accounted for, and adoption drops a catalog row that got there first.

F-Droid readiness

• Dependencies are all FOSS, from Maven Central or Google's Maven. In particular readium-lcp is deliberately absent: it pulls in the proprietary liblcp. The list users see is in LicencesScreen.kt.

• No trackers or analytics, and no Google Play services. The only outbound traffic is to the calibre-web or Komga server the user configured, and to a dictionary site when a definition is asked for. That second one is off until switched on in Settings and the site is the user's to choose (DictionaryUrl), because F-Droid review will otherwise treat a hardcoded third-party host as grounds for the TetheredNet anti-feature. Together those justify INTERNET; ACCESS_NETWORK_STATE is there for the NetworkType.CONNECTED constraint on the sync workers.

• No non-free assets. The bundled fonts (Literata, Vollkorn, Atkinson Hyperlegible, Inter) are all OFL; the icon is drawn in-repo as vector drawables.

• fonts.googleapis.com appears in the release dex and is unreachable. It is a string inside Readium's ReadiumCss, emitted only for families registered through EpubNavigatorFactory's separate googleFonts list. Liseur never sets that list: every addFontFamilyDeclaration in ReaderPreferencesMapper.kt sources its faces from bundled asset paths. Worth knowing, because a reviewer grepping the dex for hosts will find it and ask.

• Reproducible versioning: versionCode and versionName only ever change in a chore: release vX.Y.Z commit made by hack/release, and every release is tagged. F-Droid's UpdateCheckMode: Tags still notices a new tag, but AutoUpdateMode is None: under dual signing (below) every release needs its extracted signature delivered by hand, which their bot cannot do yet, so hack/release opens that merge request itself.

• Metadata lives in the repo under fastlane/metadata/android/en-US/: title, descriptions, per-versionCode changelogs, icon and screenshots.

• The build needs no network beyond Gradle dependencies and no signing config: assembleRelease on a clean checkout produces an unsigned APK, which is what F-Droid builds and signs itself.

• The build is reproducible. F-Droid rebuilds from source and will not publish a build it cannot reproduce, so hack/verify-reproducible builds the release APK twice from two clean checkouts at deliberately different paths and compares the two byte for byte:

  hack/verify-reproducible          # HEAD
  hack/verify-reproducible v0.2.0   # a tag, before submitting it
  

  If they differ it names the entries responsible, which is usually a timestamp baked into a resource or an absolute build path that leaked in. hack/release runs it on every release commit and undoes the commit rather than tag a build F-Droid could never publish; run it yourself before anything unusual ships. Move keystore.properties aside first if you have one: the check compares the unsigned APK.

  Two build-file settings exist purely to keep this true and must not be removed: dependenciesInfo is switched off (AGP would otherwise embed a dependency manifest encrypted for Google Play, which nobody else can reproduce), and packaging.jniLibs.keepDebugSymbols covers every .so (the only native code arrives prebuilt in AndroidX AARs; re-stripping it ties the bytes to the build machine's NDK: this was the one thing that made the CI APK differ from a local rebuild).

• Publishing our own signature: dual signing (merged 2026-08-21). Because the build is reproducible, F-Droid can publish the developer signature. Replacing their signature outright was declined in review, since existing F-Droid installs carry F-Droid's key and Android would refuse them every further update, so the app uses F-Droid's dual-signing flow instead: for each version F-Droid publishes two APKs, one signed with its own key (existing users keep updating, nothing breaks) and one carrying our signature, grafted onto F-Droid's own rebuild after that rebuild comes out byte-identical with ours. New installs from F-Droid get the developer-signed copy, which is interchangeable with the GitHub release. Concretely, the metadata has:

  • AllowedAPKSigningKeys: with two SHA-256 digests: our certificate and the key F-Droid signs this app with (their reviewer added the second; both APKs must pass the check).
  • AutoUpdateMode: None: their bot cannot drive this flow, so every release must arrive as a merge request adding the new build entry plus the signature files under metadata/com.chmouel.liseur/signatures/<versionCode>/, extracted from the signed release APK with fdroid signatures <apk>. hack/release does all of this: it downloads the APK from the GitHub release, extracts the signature, appends the build entry, normalises the file with fdroid rewritemeta, and opens (or, when one is already open, updates) the fdroiddata merge request. It needs fdroidserver installed (pipx install fdroidserver, or point FDROID_BIN at one); without it the release still goes out and hack/release --fdroid-only VERSION finishes the F-Droid part later.

  The merge request that set this up was fdroiddata!46390 (merged); the first verifiable tag is v0.9.3 (versionCode 17), because reproducibility landed after the v0.9.2 tag, so only 17 and later get the developer-signed twin. NewPipe (fdroiddata!46133) is the worked example this follows.

• Checking where things stand. hack/store-status prints the published versions, the index age, what the last build run did with the app, the upstream metadata, and any open merge request with its pipeline state, alongside the GitHub releases and the Play tracks, so one command answers what each of the three channels is showing.

• Submitted. The inclusion merge request was fdroiddata!44292 (merged), and hack/release opens the per-release signature merge request described above. See What F-Droid checks above for what its pipeline runs and how it can fail after a release looks finished.