The relay has always been willing to hold an inbox request until
something arrives, and the Android client has never asked it to. The
constant naming the hold was declared in MessengerService and read
nowhere, and the comment next to the 500 ms gap already described a
long poll that was not happening.
What that cost, measured against production: two devices sent 1,177
inbox requests in eight minutes, every one of them held open by the
relay for about 30 ms and answered "nothing". Every message and every
call signal waited out the gap to the next poll. Someone had already
noticed the effect on calls and shortened the gap to a second while one
was being set up, which made the polling denser without making the
signal arrive any sooner.
So `wait` reaches the wire. The signed-in mailbox is asked to hold; a
backup home never is, because two held reads in a row would put a
backup mailbox's latency in front of the one nearly all mail arrives
at. A held request needs longer than the ordinary call timeout, so it
gets a clone of the client with room for the hold.
Both loops now pace on whether the relay held the request rather than
on a fixed gap: a poll that took a while was itself the wait, and the
next one goes straight back out.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The engine has always carried compartments and both apps sealed them in,
honoured them in clearance checks and rendered them on a received file. No
screen let anybody set one, so a label was half-settable: how sensitive, but
never which part of the business. The public page promised both.
The bits travel with the file. The names do not: they are this organisation's
words, so a file from somebody else shows its areas as numbers, which is
correct rather than a gap.
Nothing is named for you. The site says areas are whatever your organisation
already calls them, so this is where you say so.
Forgetting a name does not free the bit. A file already filed under it still
carries it, and handing that bit to a different area would silently relabel
that file.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Spec §16, the third engine, checked against pinned bytes rather than against
either of the other two.
Reuses the module's own primitives: FilesKdf for PBKDF2 and Aead for GCM.
Aead returns nonce || ciphertext || tag and §16 keeps the nonce in its own
field, so the prefix comes off here; what remains is what WebCrypto calls ct.
The AAD is built by interpolation, not by a JSON writer, because it has to
match the other engines down to key order. A JSONObject would be the obvious
choice and would silently disagree, and the AAD is what stops a file being
edited into deriving a cheaper key.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two of the six suites tapped submit without filling the age gate. Submit is
disabled until birth-year holds four digits, and performClick on a disabled
Compose node does not throw: it does nothing. Each run then failed up to two
minutes later looking for a screen the tap never reached, blaming whatever
it had waited for.
A JVM test now holds it: any suite that taps submit must fill birth-year, so
the next one written cannot quietly do the same thing.
android#21 also reports that nav-contacts and nav-account no longer exist
anywhere in the app. They do. They are declared in NavDestinations.tabs and
applied through a variable, so a grep for the literal testTag("nav-" finds
nothing, which reads as the navigation having been rebuilt without them. A
third test pins that too, so the claim cannot be made again from a grep.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
None of the four repository READMEs linked to flamenetmessenger.com. The
only occurrence of the domain was inside a security contact address, which
is not a link and is not what somebody arriving at the repository needs.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
This screen showed chips and no explanation; iOS has carried one all along,
under a comment claiming both apps use the same words. They now do.
The last sentence is the boundary: a level decides who can open a file, not
where it can be put. Neither app stops somebody filing sensitive material
somewhere low, and saying so is the honest half of android#15.
Also repairs a comment the emoji sweep truncated, which left the test
saying 'the -property is NOT enforced'.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Remove dates and "this used to say X" framing from prose and comments.
What the tree does now is the only thing that matters; the history is in
git.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reported from a physical iPhone: an incoming call showed no notification
and no full-screen call.
Android never sent `urgent` on any envelope. The relay cannot work out
that an envelope is a call: it holds ciphertext, and an offer and a
birthday message are the same bytes to it. So §6.14 makes the sender
say, and only an urgent envelope gets a VoIP push. Only a VoIP push
rings a locked, suspended iPhone; anything else is a "New message"
banner, or nothing at all when the phone is locked.
So every call placed from Android reached an iPhone as an ordinary
notification, and the call screen never appeared. iOS has set this since
the flag existed (`ringsAPhone` in E2EInbox), which is why calls the
other way round work.
The detection was already here and used for the opposite question:
`warrantsAlert` identifies an OFFER to decide whether to make a sound.
Now `ringsAPhone` asks the same thing to decide whether to ring, and the
two stay separate, an answer or a hangup must not ring a phone for a
call that is already over.
4 new tests. 1221 tests pass, counted from the XML.
⚠️ Not yet observed ringing a real locked iPhone; that needs two
handsets and is the check to run before believing it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
ios#73. Android already named the reason: a call signal refused with
IdentityChangedException hung up and said "Their security code changed".
But it then told the person to go and find the conversation themselves,
and the banner that compares and accepts lives in the conversation
header, which a call started from the Calls tab or answered from a
notification never passes through. So the failure described the cure
without offering it.
The blocked peer is now carried on the phase rather than only named in
a sentence, and the failure draws a "Review their security code" button
that opens that conversation and clears the failure.
Compare and accept stay behind the existing banner. A second accept
button, phrased on a call screen in a hurry, is the thing the issue's
wording requirement exists to prevent.
4 new tests. 1217 tests pass, counted from the XML.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
My emoji sweep emptied the iOS and Android columns of every table in
this file. The check and cross marks were the values, not decoration:
this document exists to say which features exist on which client, and
with those cells blank it said nothing at all. It is also the file the
weekly parity scrub reads.
98 marks were removed by commit 4dd5440. They are back as yes, no and
partial, so the "Zero key material after use" row still records that
Android's guarantee is weaker than iOS rather than flattening it to a
tick.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every comment is reduced to its first sentence plus @param, @return
and @throws tags. Comment lines: 20040 to 8526. ktlint 1.8 applied
with its default (official Kotlin) style across all modules; 262 files
reformatted. The 264 remaining ktlint findings are naming and filename
rules, which are code changes, and are left alone.
Comments and formatting only. 1208 tests pass, counted from the XML.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
The skill's detector scored the README, PARITY.md and the extracted
comment prose. Fixed what it flagged: hollow intensifiers, "in order
to", "additionally", "the whole point", bold sentences in docs, and
bold that had wrapped across comment lines. Comment prose scores a
maximum of 4 out of 100 across the repo.
One sweep rule mangled four comments ("which is the entire point of X"
lost its clause); repaired, and the rule now keeps the clause.
Comments and strings only. 1208 tests pass, counted from the XML.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Emoji markers, markdown bold inside code comments, and em-dashes are
gone from source comments, string literals, scripts, Gradle files and
the markdown docs. Counted before and after:
emoji markers 2680 -> 1 (the emoji picker's own list, which is the feature)
em-dashes 2490 -> 9 (table placeholders and fenced code)
Two warning glyphs in user-facing strings in ChatScreens.kt are removed
as well; the sentences say the same thing without them. Two backtick
test names carried an em-dash and now carry a comma.
Nothing else changed. 1208 tests pass, counted from the result XML.
The three obfs4 interop tests are opted out with OBFS4_INTEROP_OPTOUT=1
for this run, as their own message instructs when no bridge is set up.
CONTRIBUTING gets a four-line rule so it does not grow back.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Two reports from a real phone.
**A photo auto-rotated.** BitmapFactory does not apply EXIF orientation,
and every image this app sends is re-encoded to strip EXIF — which is
right, because that block carries the GPS coordinates the shot was taken
at. The same block carries the tag saying which way is up, and a phone
camera rarely writes upright pixels: it writes landscape pixels plus
"turn this 90°". Decoding ignored that, the re-encode dropped the tag,
and the only thing that knew which way was up was gone.
The tag is now read and applied to the pixels before it is discarded, on
all three paths: preparing to send, scrubbing metadata, and rendering an
image received from a peer. The mirrored tags are handled too — a front
camera writes them, and a selfie coming back flipped is the same bug
facing sideways.
🔴 iOS is NOT affected, which is why this survived: UIImage(data:)
applies orientation as it decodes, so the identical strip-and-re-encode
is safe there. Same design, opposite platform default — and the reason
"works fine on iOS" said nothing about Android.
**The keyboard covered what was being typed into.** The manifest has
said windowSoftInputMode="adjustResize" all along, and that stopped
meaning anything at targetSdk 35: Android 15 forces apps targeting it
edge-to-edge, and an edge-to-edge window does not resize for the IME. So
the line has been a no-op on every device this ships to (minSdk 33). The
root now consumes the IME inset with imePadding() — at the root rather
than per screen, because every screen with a text field has this and one
added later would too.
5 new tests, including one that fails if the manifest line ever becomes
the only thing relied on again. 1213 tests green, counted from the
result XML rather than from BUILD SUCCESSFUL.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The Android half, matching the iOS change.
**The give-up rule (ios#22).** `flushOutbox` used
`runCatching { … }.getOrNull()`, which threw the exception away — so the loop
could not tell a contact who is merely offline from one whose identity keys
changed, and retried both for ever while the bubble went on promising
"will send". The failure is caught now and classified: a refusal from the
trust layer stops, marks the bubble FAILED and records why. Everything else
still retries for ever, deliberately — a contact who is offline or has no
device yet gets the message whenever they return.
⚠️ Classified on the exception TYPE, never on message text. A string match
would break the first time somebody reworded an error, and would break
silently: the queue would simply go back to retrying for ever.
**The reading (ios#21, #23).** `DeliveryDiagnostics` counts envelopes seen,
drops by reason, session repairs and the last send failure, per conversation.
Three or more repairs against one peer is flagged — a client that repeatedly
repairs is a symptom, not a cure.
🔴 **Shape only.** Refusal messages name what was wrong, which sometimes
means naming a device id: "device ce0c0fb4 has published keys, but none
match". Those are normalised before being counted, so the reading is safe to
screenshot and the map cannot grow a row per device id. ⚠️ Short hex ids like
`9f0e` slip past the obvious rules; the test covers them, and covers that
"base64" survives unmangled.
The new outbox test was checked by disabling the give-up rule: it fails
without it.
581 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Matches the iOS revert. Unpicks the nine commits that took Android from the
site-derived palette to purple-into-black dark-only: the purple ground, the
darkening passes, the single dark theme, the three-stop gradient, white text
boxes and white icons.
Two reconciliations, both because the surrounding work was worth keeping:
* `MainActivity`'s root `Surface` goes back to the plain one. The
transparent-Surface trap it documented -- Material deriving
`contentColor` from `contentColorFor(Transparent)`, getting Unspecified,
and falling back to BLACK text -- only existed because the Surface was
transparent to let a gradient through. No gradient, no trap.
* `CallScreen` painted `groundBrush()` to avoid stamping a flat rectangle
over that gradient. Back to `colorScheme.background`, since there is no
longer a gradient to clash with.
`PaletteMatchesIOSTest` goes with the theme it was written to hold in place.
574 tests pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
0.11 was built on 2026-09-07 and never published, so the public jumped
0.10 -> 0.12 and skipped a number. This closes that gap in the NAME, which
is the part a person reads.
The code does not go back to 11. Android refuses an APK whose versionCode
is lower than the installed one, and 12 is already published -- so code 11
would be uninstallable for anyone who took it, with no way out but an
uninstall that wipes their vault. The code is ordering; the name is the
label. They are allowed to differ, and a comment says why so it does not
get "corrected" later.
Verified: 0.11 (code 13) installs over 0.12 (code 12) with no downgrade
refusal, session intact.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A pair of translucent `primary` circles stood in for the kit's animation.
Against the gradient they read as banding -- two hard edges around the one
thing on screen -- which is what "the box around the center elements" was.
The avatar carries the centre on its own.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The send path fetches full bundles only when a device needs a session
established, so it does not drain one-time prekeys on every message. The
contact's address rides with those bundles -- so every later send had no
address, computed the `local:` key-log subject instead of `fnid:`, matched
no leaf, and refused its own send. flamenet-relay!82 puts the address on
the directory; this reads it.
Two bugs the wiring exposed:
* `rememberFnid` silently overwrote a DIFFERENT address, the one thing
its own doc comment says it must never do. An fnid is derived from the
account key, so a new address is a different account, not a new device.
It went unnoticed while only the once-per-conversation bundle path fed
it; the directory feeds it every send. It marks CHANGED now.
* `guardIdentity` read the pinned address and passed it back into the
comparison against itself, so "a different address is a different
person" could never fire on a send. It compares the served one now.
And the copy: only `IdentityChangedException` was named, so
`KeyNotLoggedException` and `KeyLogForkException` reached the screen as
"Something went wrong reaching that server." -- sending a person to check
their signal for a decision this phone had made.
Also: the call screen painted a flat `colorScheme.background` rectangle
over the app's gradient, which read as a blue box around the call area.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The receiving half already worked. `Content.Reference` was in use for
reactions, an inbound reply was recognised and rendered, and `Session.Line`
already had an `about` field to hold one. What was missing was any way to START
one: no Reply in the long-press menu, no composer state, and `Session.send` had
nowhere to put a reference.
So this is narrow, as the ticket said: `Session.send(replyTo:)` passes the
reference to `Content.text(about =)` and records it locally, a Reply item sits
above Forward in the menu, and a dismissible strip above the composer says who
is being answered.
🔴 **The preview travels with the reply, and dropping it to save bytes is not a
free optimisation.** A recipient who joined the thread later or cleared their
history has never seen the message being answered — "> [missing]" is worse than
showing what was said. A test says so rather than a comment.
⚠️ **Visible before it is sent.** A reply that only appears in the sent message
is a mode nobody can see they are in, and the commonest way to answer the wrong
message is not to know you were answering one.
⚠️ **1:1 only, and the menu item is ABSENT rather than disabled in a group.**
The group send path does not carry a reply yet; accepting one there and dropping
it would be a half-feature that looks whole. The same reasoning the file already
applies to Edit and Pin on messages written before ids existed.
Three wire tests. One of them pins that a message which is NOT a reply still
goes out as a bare body — adding this must not have turned every plain message
into an envelope an older peer reads differently.
⚠️ Not included, and its own piece: tapping the quote to jump to the message it
names. That needs a scroll-to-id in the transcript list.
158 tests, 0 failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Matches ios!146, format list included.
A photo sent through the picker had its EXIF removed — `ImagePrep` decodes to a
bitmap and re-encodes, so the output is pixels. The same photo saved into Files
and sent from the vault kept it: GPS, capture time, device serial. Somebody who
had learned "Flamenet strips location from my photos" had learned something true
of one button and false of the one beside it.
🔴 **Both doors, not one.** `sendFile` and `sendGroupFile` sit two functions
apart and had the identical gap. A leak fixed on one of two identical doors is
a leak.
🔴 **Sniffed by magic bytes, not by name.** The vault stores no MIME type, and a
file name is a guess in both directions.
⚠️ **Only images.** "Send this exact file" is what a file send means; a
scrubber that mangles somebody's PDF is worse than the leak it was fixing.
Undecodable bytes are returned as they came, for the same reason.
Five JVM tests, covering the DECISION rather than the pixels — `BitmapFactory`
needs a device, so the re-encode belongs to the instrumented suite, and iOS
covers the round trip with a real GPS-tagged JPEG. The decision is where the bug
was: the old code never asked the question at all.
⚠️ Two of the tests exist because a sloppier sniff passes the obvious ones: a
WAV file shares RIFF with WebP, and an MP4 shares the `ftyp` box with HEIC.
Falsified by matching only the RIFF header — "a WAV was mistaken for an image".
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
🔴 The accent was being used the wrong way round. It is a mid purple chosen to
be readable as a FILL with white on top of it; as the INK for a glyph on a dark
ground it is a light purple mark on a dark blue field, which reads as a
disabled control rather than a button.
`Palette.OnDark` is white. Three sets of avatar initials and fourteen
accent-tinted icons.
⚠️ The selected tab keeps the accent. Its tint reads
`if (selected) MaterialTheme.colorScheme.primary else Palette.SecondaryText` —
a conditional, so the sweep did not catch it, and that is the right outcome:
the selected tab is the one place the accent as ink is doing a job nothing else
does.
⚠️ The accent still fills: primary buttons, badges, the pill's selected state.
Ink versus fill, not "purple is gone".
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
An input is the one surface a person types into and reads back character by
character, and a dark field on a dark ground gave it the least contrast on the
screen at the moment it mattered most. Typed text reads 18.3:1 on white; the
placeholder 4.6:1, because a placeholder is still text.
⚠️ **Its own colour, not `SurfaceAlt`.** That fill is shared with incoming
bubbles and the tab pill, which stay dark — a white message bubble on this
ground would read as somebody else's app.
🔴 **Every colour a field draws has to be named, not just the container.**
Material derives label, placeholder, cursor and outline from the colour scheme,
which is the DARK scheme here: setting only `containerColor` leaves a
light-on-dark placeholder on a white box, and it simply disappears. iOS had the
same failure in a different dialect — SwiftUI drew the placeholder and caret for
a dark background until the field was given `.environment(\.colorScheme, .light)`.
Two frameworks, one mistake: a white island in a dark app has to be told it is
an island.
One `fieldColors()` in `Theme.kt`, applied to all 34 `OutlinedTextField` call
sites. Mechanical and wide rather than a new wrapper composable, because a
wrapper would have to be adopted one screen at a time and the ones that were
missed would be exactly the ones nobody looks at.
⚠️ The message composer is a `BasicTextField` and stays dark. It is a pill in
the chat, not a form field, and turning it white would put a white bar across
the bottom of every conversation.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
#3A1E9E purple, #0B3080 blue, #070A1A black, top down. Less purple than the
two-stop version, because the purple is a third of the sweep rather than half.
🔴 One card colour against three stops has to lift off all three. #0E1745
looked right against the purple and the blue and measured 1.15:1 against the
black end — a card that disappears at the bottom of a long screen. #141C52
lifts 1.41 / 1.33 / 1.23.
The palette test checks the middle stop too. A three-stop gradient checked only
at its ends is two thirds measured, and the middle is where most of a screen
actually sits.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
#3A0CA3 to #062C8C, saturation 0.93/0.96 where the old ground was 0.72/0.83.
🔴 Cards go DARKER than the ground, not lighter. Against a vivid ground a
raised card has to be lighter than a saturated violet, and secondary text then
falls below AA on it — every candidate measured did. Recessed panels invert it:
the gradient is the bright thing, content sits in dark wells.
card #190B4E (1.47:1 vs ground) · field #2A1875 · secondary #B4BAD4
avatar #31207F · danger #FF819F (was #FF6B92, 4.45:1 on the ground)
⚠️ The section band is a gap now, not a fill. It was 16pt of `surfaceVariant`,
which on the light theme was a barely-there grey seam — on the dark theme it
became a light purple slab drawing more attention than the sections it was
dividing. The panels either side of it are the separator.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Two bugs from testing on the emulator, one of them mine from this morning.
## "Erase and start fresh" erased nothing
🔴 The `VaultUnreadable` screen — shown when the keys on this phone cannot be
opened — offered exactly one action, and it called `model::signOut`. `signOut`
defaults to `erasingDevice = false`: the door that deliberately KEEPS the device
so coming back is silent. So it set the phase to SignedOut, left the unreadable
vault exactly where it was, and the next `restore()` opened it, failed again,
and came straight back to the same screen.
⚠️ It looked like it worked, because the screen changed. That is the whole
problem with a recovery action that does not recover: the feedback IS the claim,
and the person only finds out when they try to sign in and land back where they
started. Which is what happened, twice, to the person testing it.
`discardUnreadableVault()` destroys the vault, opens a fresh one, and lands on
sign-in with a sentence saying what happened and that contacts will see the
security keys change. Deliberately NOT `signOut(erasingDevice = true)`, which
would also have worked: that path first tries to revoke the device on the relay
using a session that by definition does not exist here, and its comments are
about handing a phone on. The keys are already lost; the only question is
whether the app can get past them.
Three tests, in the same source-scan shape as `DestructiveLabelsTest`. Falsified:
putting `model::signOut` back fails with the reason spelled out.
## Black text on the dark ground
🔴 `Surface(color = Color.Transparent)` with no `contentColor`. Material derives
it with `contentColorFor(color)`, which has no answer for Transparent and
returns Unspecified — so `LocalContentColor` falls back to its own default,
which is BLACK. Every `Text` that does not name a colour of its own drew black
on the dark ground.
⚠️ It read as "some of the text is hard to read" rather than as a broken theme,
because everything that DOES set a colour was fine and only the inheriting ones
went dark. That is what made it easy to ship: the screens I looked at longest
were the ones that name their colours.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Same values, same commit, as the file at the top of both palettes has always
asked for.
ground #241650 purple -> #0B2140 blue, top down
card #302F5C lifts off the ground at 1.29:1
field/bubble #3B3A6E lifts off a card at 1.20:1
ink #EDEFF9 · secondary #A8AEC9 · links #B99BFF
🔴 `onBackground`/`onSurface` are `Ink`, not `Mist`. The dark scheme used to set
them to `Mist`, which was the near-white page ground; `Mist` is the DARK ground
now, so leaving that alone would have been dark text on a dark background. That
is the exact shape of a swap that survives a rename, and it would have looked
like the app failing to draw rather than like a colour being wrong.
🔴 `MessengerTheme` no longer branches. There was a light scheme, a dark scheme,
and a stored preference choosing between them; with one palette the light branch
would draw light-mode Material chrome under dark artwork. `DisplaySettings.appearance`
is left in place so an existing stored value is not a crash — nothing reads it —
and the picker is gone from the Display section, on both platforms.
⚠️ The palette test caught one on the way in: secondary text on the avatar fill
was 4.48:1, two hundredths under. #3E3C7A became #3B3974.
The test now checks every surface text is read on — ground, card, field, avatar
— and that cards and fields visibly lift off what is behind them. A gradient
that passes and a bubble that does not is still an unreadable app.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
