Paul HittandClaude Opus 5.5
c3cd820f79
Trust.record refused unless every offered key was already pinned, so a contact who signed in on a new laptop stopped every Android user writing to them until each one tapped "Accept new keys", and their messages from the new device were held. That is the live bug; iOS has always told the cases apart. Trust.classify is the iOS table. A new key and none missing is DEVICE_ADDED: pinned at once, the verified mark cleared (the numbers compared did not cover this device), shown as a notice in the identity panel, never refused. A key missing and none new re-pins the smaller set and keeps the verified mark. A new key AND a missing one is a substitution and still CHANGED, refused until accepted, with the pinned keys kept. The receive side used to call record with the one key on the envelope, which reads as every other device removed. Trust.recordInbound takes a pinned key as it is, and for an unpinned one asks the directory (no prekey spent) and classifies the whole set it lists, exactly as a send would. A key the directory does not list, or a directory that cannot be read, is refused: unreachable is not agreement. One pinned device writing no longer shrinks the pinned set. Envelopes held while a sender was refused are replayed by the next poll once that sender is no longer CHANGED or UNKNOWN, so a contact held under the old rule for adding a device gets their messages back without a tap. A failed replay is logged and never costs the poll's other messages. The added flag is persisted in the vault's trust record (written only when set, so records from 0.31 read unchanged) and answered by the next send that finds the same keys, or by marking them verified. Tests: AnAddedDeviceIsANoticeTest (15), and TrustTest's "extra unseen key" case now expects DEVICE_ADDED, with the substitution case split out. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Flamenet Messenger. Android
The Android app, and the two libraries under it. The product's site is at https://flamenetmessenger.com.
engine/ the protocol: X3DH, Double Ratchet, sealed sender, safety numbers,
key transparency. Plain Kotlin/JVM, no Android in it at all.
client/ the relay transport: HTTP, auth, the §8 routes.
app/ the Compose app.
interop/ the cross-engine driver; not shipped.
engine and client are deliberately plain JVM libraries. The protocol has
nothing to do with Android, and keeping it off the platform means their tests
run anywhere a JVM does, no SDK, no emulator, no device, which is also what
lets them gate in CI on Linux:
./gradlew :engine:test :client:test
The app's own unit tests need the Android SDK but still no emulator (Robolectric):
./gradlew :app:testDebugUnitTest
Where it stands
PARITY.md is the definition of done for this app, read that, not this. Every row there names the iOS file it came from so it can be re-checked rather than trusted. Headlines: messaging, groups, attachments, voice notes, stories, calls, sealed sender, key transparency and the encrypted vault are built and tested, and a cross-platform round trip against iOS is verified.
If a claim here ever disagrees with PARITY.md or with the tree, believe those, and fix this.
Why the engine came first, and is held to the vectors
An Android client needs the protocol implemented a third time, and the third
implementation is the one that finds the places where the specification says
less than the two existing engines quietly agreed on, but only if it is
written against the spec and the published vectors rather than against a
reading of our own Swift and JavaScript. So the engine is held to
test/vectors/protocol-v1.json in
flamenet-e2e, the same file the
other two engines are checked against, read from that repository rather than
copied here, because a copied vector file drifts and a drifting vector file
stops being a check. PublishedVectorsTest looks for it beside this checkout
and fails loudly if it cannot find it; FLAMENET_VECTORS overrides the search.
Two implementations that agree may both be wrong in the same way, especially when one was written while reading the other. The audit-path ordering in §19 is the worked example: a top-down reading of a bottom-up path produces the correct root for every left-spine index, so it passes any test that only checks index 0. The vector file pins a right-descending index for that reason, and the Kotlin tests assert both that the published path verifies and that the same path reversed does not.
If this engine ever disagrees with the file, the interesting question is which of the three is wrong, not this one by default.
Contributing and security
- CONTRIBUTING.md, building (JDK 21,
JAVA_HOMEis not optional on a Mac), the test gate, and the platform traps that cost an hour each. - PARITY.md, what iOS does, what this does, and what is left. It is the definition of done for this app.
- SECURITY.md, threat model and where to send a vulnerability report. Not an issue tracker matter: security@flamenetmessenger.com.