Ships four app fixes merged since v1.0.14:
- #836: Generate new address now gives every enabled integration a new URL; old URLs stop working. The relay half has been live since 2026-11-19 23:39Z.
- #824: Revoke removes exactly the tapped AI client, not another client with the same name.
- #834: cert renewal waits out a rate limit instead of retrying every 15 minutes.
- #806: a missing Health Connect sends users to install it instead of crashing.
Relay-only changes (#817, #819, #820, #823) are already deployed and are not in the user-facing notes.
Owner instruction, Jason, 2026-11-19 in session: "Sure if no other fixes are on the horizon".
Reviews at b85b4e2d: konstructor 5985849806 APPROVED (outside), monkopedia-reviewer 5985849922 APPROVED.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YYSPuh9bZEJkrzvnEpY1QK
Hotfix for a 1.0.13 regression that broke every incoming connection on
Android: #781 replaced a blocking socket that ran on Dispatchers.IO with a Ktor
socket whose connect() blocks the calling thread, and the session code runs on
the main thread, so every request died with NetworkOnMainThreadException
(Crashlytics: 95 events, 3 users, first seen 1.0.13). #800 moves all of the
session's blocking work onto an IO dispatcher.
JVM tests could not see this: NetworkOnMainThreadException is Android StrictMode.
#800 adds a caller-thread test, but nothing in CI exercises the real path under
StrictMode yet.
F-Droid's bot has already queued 1.0.13 (not yet built). Tagging 1.0.14 quickly
means F-Droid offers 1.0.14 as the current version.
Owner instruction, Jason, 2026-10-22 in session: "release when ready".
Claude-Session: https://claude.ai/code/session_01YYSPuh9bZEJkrzvnEpY1QK
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Patch release so crash reports from the field become diagnosable, carrying
the app fixes that landed since 1.0.12:
- #781 (#778): a burst of abandoned connections pinned IO threads until the
whole app wedged — on the owner's phone the UI hung on its loading spinner.
Sessions moving no bytes now end after 5 minutes and hold a coroutine, not a
thread, while they wait.
- #796 (#780): a session ending on the device side now closes its mux stream,
so the client gets EOF instead of hanging and the relay frees the slot.
- #794 (#755): each release now publishes its R8 mapping.txt per flavor as a
release asset, and the relay keeps R8's map id in stack frames, so reports
from this version onward can be decoded. 1.0.9-1.0.12 reports cannot: their
mappings were never kept.
Relay-only fixes since 1.0.12 (#789, #790, #791, #793) are already deployed and
are not part of the app.
Owner instruction, Jason, 2026-10-19 in session: "when it's all landed drive a
patch release so we can get the data on this", and "Sounds fine to include
anything ready to go in."
Claude-Session: https://claude.ai/code/session_01YYSPuh9bZEJkrzvnEpY1QK
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Answers the second F-Droid source audit, run against v1.0.11 (fee84e61). That
pass confirmed all four findings from the first audit genuinely fixed, and
raised two more. Both are fixed here, plus a hang found in review.
- Refresh-token family revocation was not atomic. #750's CAS correctly stopped
the double-mint, but the CAS and the child insert were separate operations,
so a loser's deleteByFamilyId could land in the gap and the winner would then
insert a live token into an already-revoked family. 'Only one descendant'
held; 'once reuse is detected the family stays revoked' did not. Now one
transaction. (#760, #762)
- POST /token read its body with no size cap, no read timeout and no rate
limit, while /register and /mcp in the same file both wrapped their reads in
withTimeout. Now bounded on bytes READ -- not Content-Length, which a chunked
request has none of and a lying client can understate. (#761, #763)
- LiveBooleanSetting denied by suspending forever rather than returning when
the settings read failed, so four shipped MCP tools hung instead of refusing.
Fails closed either way; the refusal just never arrived. (#751, #757)
Two known test gaps are filed rather than fixed, both 'the protection works and
nothing would notice it breaking': #764 (the /token timeout has no regression
test) and #765 (the family-revocation test passes against a too-narrow
transaction -- measured, CAS-only goes green 5/5).
NOT in this release, deliberately: #754 (CT verifies the issuer allowlist, not
certificate identity) needs a design decision about renewal overlap, and #756
(three sibling ReadinessGate holders) because SecurityAlertGate's default means
'not alerting', so the identical patch would fail OPEN there.
Claude-Session: https://claude.ai/code/session_01YYSPuh9bZEJkrzvnEpY1QK
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Answers the 2026-09-14 source audit on MR !42096. Everything below is already
on main; this pins it to a tag F-Droid can point at.
- The "Allow AI to act on notifications" and "Allow DND changes" switches
were stored but never consulted by the production providers, so both tool
families stayed live with the switch off. On 1.0.10 set_dnd_state returned
success and actually silenced the phone. (#746, #748)
- HttpHeaderInjector buffered the header block and chunk-size line with no cap
ahead of Ktor, so Ktor's limits never applied. Measured: 16,793,629 bytes
retained, nothing emitted, no exception. (#749, #750)
- Refresh-token rotation was read -> check -> write with nothing holding the
gap; 199 of 200 rounds minted two descendants. Now a single CAS UPDATE.
Also carries the listing and privacy corrections promised on 09-10 but landed
after the 1.0.10 tag, so the pinned tree never contained them: full_description
said the certificate check "Runs on an interval you pick, or not at all" when
it ships on at 12 hours. (#743, #744)
Claude-Session: https://claude.ai/code/session_01YYSPuh9bZEJkrzvnEpY1QK
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Android 16 (API 36) is the target level Play will require for new
uploads. `app/build.gradle.kts` holds the only `targetSdk` declaration
in the tree -- the other four Android modules declare `compileSdk` and
`minSdk` but inherit the target -- so this is a one-line change plus the
two places that record the value as prose: the F-Droid maintainer note
and the README build prerequisites.
`compileSdk` was already 36 everywhere, so nothing about the compile or
dependency surface moves. The four `robolectric.properties` files stay
pinned at `sdk=33` on purpose, which is why all 179 goldens are
unaffected: they render at 33 regardless of what the app targets.
No version bump and no changelog entry -- versionCode/versionName stay
at 11/1.0.10. This rides whatever release the owner cuts next; a
changelog for a versionCode that does not exist yet would confuse the
release gate.
Claude-Session: https://claude.ai/code/session_01YYSPuh9bZEJkrzvnEpY1QK
Co-authored-by: Jason Monk <monkopedia@gmail.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* Bump to v1.0.10
Two user-facing changes since 1.0.9, both already on main:
- Crash reporting in the FOSS build is opt-in and off by default, with a
first-run consent sheet and a switch in Settings > Support (#546, #735).
The google/Play distribution is deliberately unchanged.
- The security-check interval gains `Never`, which stops the periodic work
AND the opportunistic run TunnelForegroundService enqueues on connect --
the only two callers of the crt.sh / Certspotter fetchers (#740, #741).
changelogs/11.txt already existed (written by #735) and covered only the
first. Rewritten to cover both, 493 bytes -- under the 500-char cap fdroid
lint enforces. Verified against the built APK rather than asserted: an
assembleRelease reports versionCode 11 in output-metadata.json, and the
release.yml gate replayed against it passes with 405 non-whitespace bytes,
so #607's whitespace loophole is not what is carrying it.
The `Never` option makes a scheduled outbound request optional rather than
absent, so two claims that read as absolute needed adjusting:
- fastlane full_description.txt gains a Certificate monitoring bullet. The
existing "Private by design" bullet is left as-is, deliberately.
- docs/user/privacy.md dropped "no background upload" -- a scheduled CT
query contradicts it. Replaced with the narrower true claim (no analytics,
no telemetry, nothing about you uploaded) plus a paragraph naming crt.sh /
Certspotter and the Settings > Security control. README.md and the welcome
string are untouched: the README claim is about user data and stays true,
and the welcome string is a tagline, not a claim.
Two changes to fdroid/com.rousecontext.yml beyond the version fields:
- An app-level AntiFeatures: TetheredNet block. BuildConfig.RELAY_HOST is
fixed at build time and F-Droid builds bare, so the shipped APK talks to
relay.rousecontext.com only. Declared rather than left for a reviewer to
find. Placement and indentation checked against a live fdroiddata recipe
(app.myzel394.locus.yml), not inferred.
- The `Description:` block is removed. The submitted copy (fdroiddata!42096)
has no Description -- the maintainer asked for HTML, and fastlane's
full_description.txt supplies it. Keeping a second copy of the prose in
the template is how the listing drifts from the app, which is the shape of
#737. A comment records where the text now lives.
Stale `v1.0.8` references in the Builds comments (tags that exist now)
refreshed to v1.0.10 while in there.
No tag is pushed. Tagging triggers release.yml and publishes signed
artifacts; that is the owner's separate step after this merges.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YYSPuh9bZEJkrzvnEpY1QK
* Close the second exhaustive-network claim in privacy.md
The brief named one line to fix, but "The only network calls are to the relay
... and to Firebase Cloud Messaging" is the same claim in a different section,
and it is false in exactly the same way: it enumerates outbound destinations
exhaustively and omits crt.sh / Certspotter. Fixing the first line and leaving
this one would have left the doc contradicting itself two sections apart.
Now non-exhaustive and conditional ("while certificate monitoring is switched
on"), pointing back at the paragraph that names the hosts rather than
repeating them.
Scoped to the CT claim only. This section also says FCM where the FOSS build
uses UnifiedPush -- a real, separate divergence that predates this release and
is not what this bump changed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YYSPuh9bZEJkrzvnEpY1QK
* Pin the version in screenshot fixtures so goldens survive a release
The 1.0.10 bump turned four goldens red: settingsCrashReports{On,Off}{Light,Dark}
(12 failures = 4 goldens x 3 retries). SettingsState.versionName defaults to
BuildConfig.VERSION_NAME, SettingsScreen renders it in the About row, and those
four captures scroll past it. "Version 1.0.9" became "Version 1.0.10" and the
pixels moved. Measured, not inferred: the golden-vs-actual diff is a single
320x33 band at the version row in all four, and nothing else.
This is the first bump since those goldens existed -- #735 merged the day
before -- which is why 1.0.9 did not hit it.
Re-recording alone would have been the wrong fix. It clears the board and
re-arms the trap: a golden that embeds the version number fails once per
release, forever, and whoever hits it next gets no signal that it is expected.
So the version is pinned to a synthetic literal instead, the same way the test
JVM pins user.timezone=UTC (#633) so goldens do not depend on the recorder's
environment. BuildConfig.VERSION_NAME is environment in exactly that sense.
Pinned at all eight SettingsState fixtures in the screenshots package, not just
the four that move today, so a capture that later scrolls further is already
safe. The crash-report fixture moves to a file-level `internal fun` so the
guard can assert against the real fixture rather than a copy of it.
The default itself is deliberately NOT changed. Production never assigns
versionName -- the default is the only thing that puts a version on the About
row in the shipped app -- so changing it would blank a real screen to fix a
test. A third assertion holds that default in place.
ScreenshotVersionPinTest carries the guard, and it is not vacuous. Negative
control: dropping `versionName = SCREENSHOT_VERSION_NAME` from the fixture
fails it with "expected 0.0.0-screenshot but was 1.0.10" and a message naming
the goldens at risk. It also asserts the pin is never equal to
BuildConfig.VERSION_NAME, which is the assertion that survives the tempting
wrong fix -- setting the pin to the current release looks like a pin, passes
that day, and silently restores the breakage.
Proof the pin works, not just that today is green: with versionName temporarily
set to 99.99.99-pinproof, :app:verifyRoborazziDebug passes and
check-roborazzi-coverage.sh reports 139 distinct goldens compared, so the green
is a real comparison rather than the compared-zero state #628 describes.
Only four PNGs are rewritten, each differing from its committed version by that
one band and nothing else. Golden counts are unchanged: 139 app, 40
notifications.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YYSPuh9bZEJkrzvnEpY1QK
---------
Co-authored-by: Monkopedia <monkopedia@gmail.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Carries the ANR fix that has been looping on a real device since
2026-06-14, plus a run of tunnel-reliability fixes.
User-facing:
- ACRA's logcat collector no longer hangs the app. `withLogcatArguments(emptyList())`
did not disable logcat collection — it removed the only flag that made the
subprocess terminate, so ACRA ran bare `logcat`, which streams forever, and the
caller blocked on a no-timeout FutureTask.get(). Measured on-device as an ANR
every ~65s in a kill/restart loop (#583).
- ACRA silent reports dispatch off the caller's thread (#542).
- Inbound mux frames are demultiplexed off Dispatchers.Main (#569).
- A TLS record spanning two mux DATA frames now completes the handshake (#563).
- Inbound mux frames are serialised so DATA keeps wire order (#562).
- The tunnel tears down when the inbound frame consumer throws, instead of going
silently deaf (#568).
Also: FOSS purity is now verified against the built APK in CI (#548), and three
CI gates that were silently swallowing real violations are anchored and
self-tested (#577, #590, #594).
F-Droid recipe moved to 1.0.9 / versionCode 10. Gradle wrapper stays at 8.14.3 —
gradlew-fdroid has no hash for 8.14.4, and AGP 8.13.2 caps us below Gradle 9.6.0.
* ci: verify FOSS purity against the built APK (#548)
The FOSS/Google split had no verifier. `project.hasProperty("google")` is
satisfied by any Gradle property source — including an ambient
`ORG_GRADLE_PROJECT_google` — and nothing downstream ever looked at the
artifact, so a Firebase-carrying build published as
`rouse-context-<v>-foss.apk` would have shipped silently.
Add scripts/check-apk-distribution.sh, which unzips an APK and counts
com.google.firebase / com.google.android.gms references plus the
google-services-generated string resources, then asserts presence or absence
against the distribution the artifact claims to be.
It refuses to report a pass it did not earn: a missing/small APK, an unzip
with no entries, no classes*.dex, a grep error, a pattern that no longer
matches its synthetic fixture, or an APK with no com/rousecontext references
all fail loudly rather than reading as clean.
Wired both directions. ci.yml checks each debug APK immediately after its
build (both variants share one output path). release.yml checks the STAGED
dist/ files — step ordering plus a copy is the only thing distinguishing the
two release artifacts. The --expect google direction is the live positive
control: it fails if the patterns ever stop matching real Firebase code.
Also reject the two property sources nobody can see from the build command
(ORG_GRADLE_PROJECT_google and the gradle-user-home gradle.properties), log
the resolved distribution at configuration time, and correct the false comment
that claimed the google-services plugin's hard-fail-on-missing-credentials
protected the gate — google-services.json is committed, so there is nothing
for it to fail on.
* ci: make the hit-collection branch explicit in the purity gate
`[ "$n" -gt 0 ] && hits+=(...)` is exempt from `set -e` (it is the command
before the final &&), so it worked -- but a reader has to know that rule to see
it. Spell it as an if.
* ci: scope the purity patterns to DEX class refs vs everything else
Measured against real APKs rather than assumed. A whole-APK
`com.google.android.gms` grep is 0 on the minified FOSS release APK but 4 on
the FOSS debug APK, and all four are legitimate: they are Intent-action and
security-provider NAME STRINGS that androidx.activity and Conscrypt carry
whether or not Play Services exists (…gms.provider.action.PICK_IMAGES,
…gms.org.conscrypt, and two more). No GMS code is linked. ci.yml builds debug
APKs, so the first draft of this gate would have been permanently red.
The discriminator is the separator. A DEX type descriptor uses slashes, so the
DEX scope requires the slash form: FOSS 0/0, Google 2869/6571. Outside the DEX
— manifest, resources.arsc, res/, assets — either separator counts, because
that is where component class names and the google-services-generated
google_app_id / gcm_defaultSenderId / firebase_database_url resources live, and
the FOSS build has genuinely zero of them.
Both scopes keep their own synthetic fixture, so neither can rot unnoticed.
Also make the corrected build.gradle.kts comment precise about what was
measured: the google-services plugin's hard-fail on missing credentials is
real (verified — moving google-services.json aside fails
:app:processDebugGoogleServices), it just protects nothing here, because the
file is committed and the hard-fail therefore never fires.
---------
Co-authored-by: monkopedia-coder <monkopedia-coder@users.noreply.github.com>
Carries #538 (health reads paginate + bucket/downsample, fixes#536) and #543
(HTML Fastlane description) into a release, so the F-Droid recipe can point
`commit:` at a real release tag instead of a side commit off v1.0.7.
Adds the versionCode 9 changelog and syncs fdroid/com.rousecontext.yml to
1.0.8/vc9. Also records that the submitted fdroiddata copy carries no
MaintainerNotes field, at the maintainer's request on fdroiddata!42096.
Claude-Session: https://claude.ai/code/session_01YYSPuh9bZEJkrzvnEpY1QK
Co-authored-by: Monkopedia <monkopedia@gmail.com>
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
The staged recipe's prebuild had the first-draft path (rm -f gradle/...) which
is wrong: prebuild runs in the app/ subdir, so it needs ../gradle/..., and it
must also strip the two apply-false Google/Firebase plugin declarations from the
root build.gradle.kts (the scanner flags them otherwise). This matches the
fdroiddata MR !42096 recipe, which was verified with a full fdroid build
(scanner clean + reproduces the published FOSS asset).
Co-authored-by: Monkopedia <monkopedia@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
The FOSS build (bare `assembleRelease`) already skipped applying the Firebase
Gradle plugins at runtime, but every Firebase/Google-services coordinate — the
two plugin ids, the Crashlytics per-build-type config, and the firebase-*
dependencies — still lived inline in app/build.gradle.kts, guarded only by
`if (googleBuild)`. F-Droid's source scanner is static, so it still saw those
non-FOSS coordinates in a file that survives into the FOSS build. The fdroiddata
maintainer asked us to remove the non-FOSS deps in prebuild rather than
allow-list them with scanignore.
This extracts ALL Google/Firebase build config into
gradle/google-services.gradle.kts, applied from app/build.gradle.kts ONLY under
`-Pgoogle` (`if (googleBuild) { apply(from = ...) }`). The bare FOSS build never
applies it, so it compiles ZERO Firebase: no plugins on the classpath, no
firebase-* deps, no Crashlytics config, and google-services.json is not required
to assemble it. Verified: aapt2 dump strings on the FOSS APK shows no
firebase/gms strings and the DEX contains no com.google.firebase /
com.google.android.gms classes, while the `-Pgoogle` APK still links Firebase.
The extracted file lives in the repo-root gradle/ dir (not app/) because Android
Lint's build-script visitor (lintVitalRelease, which the FOSS release build runs)
crashes trying to FIR-resolve a standalone apply(from=...) script inside an
Android module; the root project has no Android lint task. Its Crashlytics
mapping/symbol toggles are set by reflection because an apply-script has no AGP /
crashlytics DSL types on its classpath.
The fdroiddata recipe now removes the file in `prebuild:
rm -f gradle/google-services.gradle.kts` before the scanner runs (no
scanignore/scandelete), and the version/toolchain fields + MaintainerNotes are
updated to 1.0.7. Adds changelog 8.txt (build-internal, no user-facing change
vs 1.0.6). Part of #465.
Co-authored-by: Jason Monk <monkopedia@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
F-Droid's gradlew-fdroid verifies the Gradle wrapper checksum against its
own hash list, which tops out at 8.14.3 for the 8.14.x line (it skipped
8.14.4 and moved to 9.x). The project pinned 8.14.4, so an F-Droid
reproducible build fails hard with "No hash for gradle version 8.14.4!".
Confirmed by running fdroid build locally.
Pin the wrapper to 8.14.3 (highest 8.14.x F-Droid verifies, one patch below
current, fully compatible with AGP 8.13.2) and bump to 1.0.6 (versionCode 7)
so the fix ships in a tracked release. Update the fdroid recipe, changelog,
and docs accordingly; 1.0.6 is now the first version submitted to F-Droid.
Part of #465.
Co-authored-by: Monkopedia <monkopedia@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
v1.0.4 was tagged before the FOSS wake-banner fix (#531) merged, so the
first actual F-Droid debut should be 1.0.5. Current main is v1.0.4 + #531
(pending-setup vs needs-setup Home banner) + the #529 fdroid recipe key-fix.
- app/build.gradle.kts: versionCode 5->6, versionName 1.0.4->1.0.5
- fdroid/com.rousecontext.yml: Builds entry, CurrentVersion(Code), and the
MaintainerNotes toolchain label to 1.0.5 / v1.0.5 / 6 (historical
key-confirmation refs to v1.0.4 kept; signing key and Binaries untouched)
- fastlane changelogs/6.txt: new user-facing changelog owning the
"first F-Droid release" framing; 5.txt softened since it never shipped
- docs/distribution.md: forward-looking release-to-cut ref 1.0.4->1.0.5
Part of #465
Co-authored-by: Jason Monk <monkopedia@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Published FOSS releases (v1.0.2, v1.0.3, v1.0.4) are signed by GitHub
Actions using secrets.RELEASE_KEYSTORE, whose certificate SHA-256 is
7cc8d2d568eb3d20a5e190e77baa97b3bde80782dd2576f29088a16c4ce47850
(verified via `apksigner verify --print-certs` on the published
rouse-context-1.0.4-foss.apk). The recipe previously declared the local
.signing/release.keystore cert (5394b7c7...), which does NOT match any
published release and would cause F-Droid's AllowedAPKSigningKeys check
to reject the upstream-signed asset.
Update the YAML value, the MaintainerNotes describing it, and the
matching docs/distribution.md reference so none of them tie the value to
the local keystore. Owner confirmed the CI key is canonical.
Part of #465.
Co-authored-by: Jason Monk <monkopedia@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Bump versionName 1.0.3 -> 1.0.4 and versionCode 4 -> 5. The current v1.0.3 tag
predates all the F-Droid groundwork (reproducible-build config, recipe, fastlane
metadata) and the #527 header fix, and its published release has no FOSS-named
APK, so F-Droid must build a fresh release from current main.
- app/build.gradle.kts: versionCode 5, versionName 1.0.4.
- fdroid/com.rousecontext.yml: Builds entry + CurrentVersion(Code) to 1.0.4/5;
refresh comments (v1.0.4 framing, instruct fdroiddata copy to pin commit: to
the immutable release SHA, cite healthdisconnect as live-on-F-Droid proof of
the reproducible-Binaries path, document the load-bearing reproducibility pins
and the bare-assembleRelease requirement).
- fastlane changelogs/5.txt: first real F-Droid release notes; soften 4.txt's
"First F-Droid release" line since 4 never shipped.
- docs/distribution.md: note the reproducible path is proven in production by
com.monkopedia.healthdisconnect.
Part of #465.
Co-authored-by: Jason Monk <monkopedia@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Make the FOSS release build buildable by F-Droid from a clean checkout
with no signing credentials, mirroring com.monkopedia.healthdisconnect's
proven reproducible-builds setup (live on F-Droid).
- Remove the hard-fail signing guard (gradle.taskGraph.whenReady). The
owner has decided to match health-disconnect, which has no such guard:
absent credentials must produce an UNSIGNED build, not an abort. On AGP
8.13.2 an incomplete-but-assigned signingConfig still aborts packaging
(health-disconnect's older AGP silently skipped it), so the release
signingConfig is now assigned to the buildType only when the keystore
and both passwords are present (canSignRelease). Credential-less
assembleRelease builds unsigned; a build with creds still signs.
- dependenciesInfo { includeInApk/InBundle = false } — AGP's embedded
dependency-metadata block is rejected by F-Droid's scanner.
- vcsInfo.include = false on release — embedded VCS/commit info is
non-reproducible.
- Disable ArtProfile tasks — Compose pulls in profileinstaller, whose
baseline.prof generation is non-deterministic and breaks reproducible
builds.
- Simplify the staged fdroiddata recipe: drop the throwaway-keystore
prebuild block and its MaintainerNotes paragraph (unnecessary now the
bare build builds unsigned). Recipe stays an owner-gated DRAFT.
Verified both ways: credential-less :app:assembleRelease produces
app-release-unsigned.apk (apksigner: DOES NOT VERIFY); with creds it
produces a signed app-release.apk (cert SHA-256 5394b7c7…51dca). Both
FOSS and -Pgoogle variants compile.
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Author DRAFT groundwork for the F-Droid (fdroiddata) submission, staged in
this repo only — NOT submitted upstream. Submission stays an owner-gated
manual step.
- fdroid/com.rousecontext.yml: draft fdroiddata metadata using the
reproducible-builds path (Binaries: GitHub Release foss APK template +
AllowedAPKSigningKeys: release cert SHA-256), a Builds entry for the
no-flag FOSS default variant (subdir app, gradle [yes]) pinned to the v%v
release tag, throwaway-signing prebuild so a credential-less source build
completes, and CurrentVersion(Code) referencing app/build.gradle.kts.
- docs/distribution.md: which variant ships where (Google = sideload APK /
Play AAB; FOSS = F-Droid + GitHub foss APK), the UnifiedPush-distributor
first-run requirement, the single shared relay backend (per-device
push/auth discriminator), and the signature story (reproducible build →
F-Droid ships our signature → FOSS APK ↔ F-Droid with no uninstall, but
Google and FOSS are different identities — don't cross).
- README + self-hosting.md: one-line pointers to the new doc.
Co-authored-by: monkopedia-coder <monkopedia+coder@gmail.com>
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>