- Account.kt: collapse three identical backend-not-configured Failure
constructions into one helper
- OnchainZapSendError: declare causeIsUserFacing on the enum so the
sender owns which failures carry a human-readable cause, instead of
the UI mapper hardcoding the list
- SendPaymentScreen: fee chip label is now a single format resource
instead of manual string concatenation
Surfaces relays unresponsive for 7+ days across the user's NIP-65 (10002),
DM (10050), and Search (10007) relay lists. A non-modal banner appears
above feed columns (and above the single-pane content) whenever the
classifier finds anything; tapping it opens an anchored Popup with one
row per unhealthy relay and per-row Remove / Open Dashboard / Snooze 7d
actions plus a banner-level "Snooze all 7d".
Quartz
- RelayStat gains best-effort lastConnectAt + lastIncomingAt timestamps
(epoch seconds, 0 = never observed). RelayStats listener pushes them
on onConnected / onIncomingMessage. Durable per-relay history lives
outside quartz in the commons RelayHealthStore.
Commons (new commons/relays/health/ package)
- classifyRelayHealth() pure function with the v1 gates:
* first-run grace (don't flag for 7d after firstScanAt)
* offline grace (don't flag if no relay anywhere has responded)
* Tor-mode skip (relay timing is intentionally lossy through Tor)
* per-relay snooze (snoozedUntil > now)
* 10006 (blocked) excluded from detection but still part of the
multi-list Remove action
- RelayHealthStore (account-scoped, supervised scope, 5s debounced
persist, 60s ticker for snooze expiry).
- RelayHealthListener wires the quartz lifecycle into the store.
- RelayHealthPersistence interface (no expect/actual — single impl per
platform via injection).
- RelayListMutator interface + RelayRemovalResult sealed type.
- Shared UnhealthyRelayBanner (errorContainer @ 50% alpha) and
UnhealthyRelayRow (static outlined tag chips, no ripple) composables.
- 8 classifier unit tests covering each gate + multi-list membership.
Desktop wiring
- PreferencesRelayHealthPersistence (java.util.prefs.Preferences, per
account via 8-char pubkey prefix).
- DesktopRelayListMutator runs the 4 sign-and-broadcast jobs in
parallel via async/awaitAll so a slow NIP-46 bunker doesn't multiply
latency by 4.
- Banner placed in DeckColumnContainer + SinglePaneLayout, store +
listener + per-account scan trigger wired in Main.kt's MainContent.
Scope: Desktop only for v1. Android wiring is intentionally not in
this PR — the commons module is platform-neutral and ready for Android
to follow whenever someone wants to pick it up.
Quartz: new experimental/fitness/workout package shaped like nip88Polls —
WorkoutRecordEvent with per-tag classes (exercise, duration, distance,
elevation, calories, steps, heart rate, splits, strength sets/reps/weight,
source, workout_start_time), TagArrayBuilder/TagArray extensions, lax
RUNSTR-dialect parsing (unit defaults, HH:MM:SS or raw seconds), and
EventFactory + LocalCache registration. Covered by fixture tests.
Amethyst: new Workouts feed (drawer entry, route, follow-list top bar,
per-relay filter assemblers mirroring the Pictures feed) with a + FAB
opening a manual workout composer that publishes canonical kind-1301
events. Workout cards render stats chips and also display inside threads
via NoteCompose. Adds fitness Material Symbols glyphs and regenerates the
subset font.
https://claude.ai/code/session_01Kpx53UEeJqqR7CASzMu6GB
Lightning payment targets already route into the Send Payment screen;
this extends the same treatment to bitcoin targets. Tapping a profile's
bitcoin payment-target chip (or its pay action in the wallet-button
dialog) now opens the Send Payment screen with the on-chain rail locked
to that announced address, paid directly from the user's NIP-BC Taproot
wallet — falling back to the external bitcoin: URI when the chain
backend is missing or the address isn't a payable native-segwit mainnet
address.
- quartz: SegwitAddress.scriptPubKeyFor/isPayableMainnetAddress;
OnchainZapBuilder.buildToScripts core shared by the pubkey paths.
- commons: OnchainZapSender.sendToAddress — plain wallet send with the
same fund-safety signing contract but no kind:8333 receipt (the
destination isn't pubkey-derived, so none is possible); the signing
block is now a single shared helper across send/sendSplit/sendToAddress
and Success.receiptEventId is nullable for receipt-less sends.
- amethyst: Account.sendOnchainToAddress; Route.SendPayment gains
btcAddressOverride; a shared inAppPaymentRouteFor() decides which
payment targets the user's wallets can pay in-app (used by both the
target chips and the payment-targets dialog).
- Send Payment screen: with an address override the on-chain rail shows
the target address, hides the message field (no receipt to carry it),
explains that no zap receipt is published, and dispatches the plain
address send.
https://claude.ai/code/session_01UERRsbDoRPz46Qx5HCXgAa
- Clamp the seeded kind:445 subscription since at wall-clock now: the
inner createdAt is sender-controlled, so a single future-dated message
could push since past the present and silently skip genuinely new
events on every restart. Covered by a new regression test.
- Drop the remember() around the group-list unread count: the chatroom's
message set can shrink without newestMessage or lastReadTime changing
(pruning, kind:5 deletion of an older message), which left the cached
count stale. The set is pruned to ~100 entries, so counting per
recomposition is cheap.
- Extract marmotGroupLastReadRoute(): the "MarmotGroup/<id>" last-read
key was inlined at three call sites; a prefix drift between the
mark-as-read side and the unread checks would silently reintroduce
the bug this branch fixes.
- Derive GROUP_EVENT_REFETCH_OVERLAP_SEC from TimeUtils.ONE_DAY instead
of re-deriving 24*60*60.
Relay-socket Tor routing only had localhost/onion/DM/trusted/new buckets,
so a wallet or payment-service relay fell through to newRelaysViaTor and
got forced over Tor regardless of the "Money operations via Tor" toggle
(which previously governed only HTTP clients). On services that block Tor
exits this silently broke NIP-47 and CLINK payments.
Add a moneyOperationsViaTor field to TorRelaySettings and a moneyOpRelay
bucket to TorRelayEvaluation (taking precedence over DM/trusted/new, after
the onion reachability check). TorRelayState gains a persistent money-op
relay set — fed across all accounts from NIP-47 wallet relays and saved
CLINK debit relays via AccountsTorStateConnector — plus a reference-counted
ad-hoc registry for one-off payment relays (e.g. an noffer pointer). The
websocket builder resolves the per-relay decision from live source values
so ad-hoc registration takes effect on the next connect with no race.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The test lived in commons androidHostTest, but no CI workflow or
pre-push task runs :commons:testAndroidHostTest — and running it
manually fails before reaching any assertion: quartz's android
PlatformLog actual hits unmocked android.util.Log stubs
(NoSuchMethodError), since the source set is not configured with
returnDefaultValues. The end-to-end leave/rejoin coverage was
therefore never executed anywhere.
:commons:jvmTest runs in CI and in the pre-push hook, already has the
secp256k1 JVM bindings the test needs, and uses quartz's JVM logger.
Verified green there alongside MarmotManagerRestoreTest.
The Marmot subscription since, the processed-event dedup set, and the
application ratchet position (group state persists only at commits) are
all in-memory only. On restart, relays therefore redeliver the group's
entire kind:445 history and the rewound ratchet re-decrypts old
application messages as if they had just arrived — wasted decryption
work and, when a replay beats the disk restore, duplicate entries
appended to the persisted plaintext message log.
Two defenses:
- MarmotManager.restoreAll() now seeds each restored group's
subscription since from the newest persisted decrypted message, minus
a one-day overlap window for late/out-of-order publishes. Seeding
happens before syncWithGroupManager registers default entries, so
even the first filter set sent to relays carries it. The CLI is
unaffected: it builds group filters from its own persisted since.
- MarmotMessageStore appends are now explicitly idempotent (contract
was previously ambiguous and both real stores appended blindly):
the Android and CLI file stores skip an entry that is already in the
group's log, so replays inside the overlap window cannot grow it.
Covered by MarmotManagerRestoreTest in commons jvmTest — placed there
rather than androidHostTest because CI only runs :commons:jvmTest (the
androidHostTest task currently fails on android.util.Log stubs even
for the pre-existing Marmot test).
The unread dot for Marmot/MLS group rooms was driven by an in-memory
unreadCount on MarmotGroupChatroom. On restart the MLS group state is
restored from the last persisted commit, the kind:445 subscription
restarts with since=null, and the in-memory processed-event dedup set is
empty — so relays redeliver old group events, they re-decrypt as fresh
application messages, and the counter was re-bumped, resurrecting the
dot for chats already read.
Marking-as-read was already persisted: opening a group chat writes the
newest rendered message's createdAt to the MarmotGroup/<groupId> route
in lastReadPerRoute (saved to disk with account settings). Compute the
unread indicators from that timestamp instead — exactly how DM rooms
and public channels do it — in both the Messages screen row and the
Marmot group list row.
With no consumer left, drop the volatile counter and collapse the
addMessageSync/restoreMessageSync split (they only differed in the
counter bump).
Brings the MLS/Marmot message composer to parity with NIP-17 DMs:
typing @ shows the shared user-suggestion dropdown (local cache +
NIP-05 resolution), selecting a user inserts @npub…, and on send
NewMessageTagger rewrites mentions into nostr: URIs and collects
the referenced users as p-tags on the inner kind:9 rumor. Mentions
stay inside the MLS ciphertext; the outer kind:445 is unchanged.
Also applies MentionPreservingInputTransformation and
UrlUserTagOutputTransformation to the field so mentions render
highlighted while composing, matching the DM editor.
https://claude.ai/code/session_013NWdjCSegsf2FYSPPANX3n
From a spec/SDK audit (verified against the CLINK spec, not just SDK 1.5.5):
- NOffer.price: decode as UNSIGNED 4-byte big-endian (now Long) — the SDK reads
price via parseInt(hex); reading it signed turned prices >= 2^31 sats negative
and broke encode/decode idempotency for high-bit prices.
- Manage (21003) messages corrected to the nested spec shape: request nests offer
data under offer{id,fields}, payer_data is a string list (not a map), and the
response uses details + field (was offer/offers). Documented the single-object
details limitation (Manage is consume-unused).
- DisplayClinkOffer: cache NIP-05 .well-known clink_offer lookups (incl. negative
results) so profile visits / kind-0 refreshes don't refetch nostr.json.
Deliberately NOT changed: the offer 'latest' (code 3) field and ndebit k1 at
TLV-3 — both are SPEC-defined; the SDK 1.5.5 merely lags, as the code comments
already noted. CLINK tests pass; app compiles.
fixMissingSpaces runs on the main thread once per rendered note. The scan
introduced in the previous commit re-tested every detected URL at every
character position (and allocated an iterator per position via firstOrNull),
i.e. O(N*U*L) for a note with U URLs — noticeable on large notes that carry
many links.
Bucket the URLs by their first character once up front and only attempt a
match at positions whose character can actually start a URL; every other
character now costs a single map lookup, keeping the pass linear in the text
length for typical content. Buckets stay longest-first so prefix URLs still
don't shadow longer ones, so the output is identical (verified against the
commons richtext JVM corpus).
RichTextParser.fixMissingSpaces used a Regex of the form
`([^ \n])?(urls)([^ \n])?` to insert spaces around URLs glued to
neighbouring text. Kotlin/Native's regex engine fails to backtrack the
optional `([^ \n])?` capture groups to zero width, so on iOS every URL was
corrupted (e.g. "https://x" became "h https://x"). That broke the
downstream segmenter, which is why :commons:iosSimulatorArm64Test reported
19 failures across the RichText/Gallery/Pdf/F4a parsers once the test binary
finally linked.
Replace the regex with a direct left-to-right scan that inserts a single
space wherever a detected URL touches a non-space/non-newline neighbour. The
scan is engine-independent, so it behaves identically on JVM and Native.
Verified equivalent to the old behaviour across the full commons richtext
JVM corpus, and the new FixMissingSpacesTest pins the cases on every target
(including iosSimulatorArm64).
Adds the verifiable core for using a CLINK debit pointer as a spend rail
alongside NWC:
- ClinkDebitWalletEntry (commons): a saved ndebit pointer, the spend-only
counterpart of NwcWalletEntry (no secret, no balance/history)
- PaymentSource + PaymentSourceResolver (commons): unifies NWC wallets and
CLINK debits into one list with a single default id spanning both types;
no explicit default falls back to first (NWC before debits), preserving
today's behavior. canShowBalance marks NWC vs debit honestly.
- ClinkDebitPayer (amethyst): publishes the kind-21002 pay request and awaits
the preimage via a one-shot subscription, mirroring ClinkOfferPayer.
Resolver logic covered by PaymentSourceResolverTest on JVM (7 cases incl.
cross-type default + stale-id fallback); amethyst compiles. Persisting the
new fields in AccountSettings and the Wallet-screen rows/confirm dialog are
the next (compile-only) step.
The :commons:linkDebugTestIosSimulatorArm64 CI step fails under Xcode 16.4
with 'Undefined symbols: _OBJC_CLASS_$_UIViewLayoutRegion'. The symbol is
referenced by the prebuilt Kotlin/Native cache of Compose Multiplatform's
org.jetbrains.compose.ui:ui-uikit (CMPLayoutRegion), which was built against
a newer simulator SDK (18.5) than the test binary is linked for (14.0).
Disable the native compiler cache for the iOS test binaries so ui-uikit
recompiles against the active SDK, where UIViewLayoutRegion resolves. The
DisableCacheInKotlinVersion guard re-surfaces the workaround once we move
past Kotlin 2.3.21 so it can be removed when the cache is fixed upstream.
Teaches the commons RichTextParser to recognize an inline noffer1...
token and emit a ClinkOfferSegment carrying the decoded NOffer, so a
GUI front end can render a 'Pay' card in the note body (the feed-offer
feature). Bare tokens only for now; nostr:/lightning: prefixed forms
fall through. Covered by ClinkOfferSegmentTest on JVM.
Backs out the reactive list-refresh plumbing added in the prior commit:
removes the `changes` SharedFlow from the shared `ChatroomList` (restoring
it to its original form) and reverts Desktop `ChatroomListState` to its
original 2s poll. Room assembly is expected to move to a LocalCache.observe
approach on both platforms later, which would supersede this.
Keeps the independent Desktop list improvements (per-room unread tracking
and the mute/acceptable filter), which don't depend on the flow.
https://claude.ai/code/session_01VEukNczAYxNLBjLnqVEoZd
Makes the Desktop DM client a first-class group participant and tightens
the shared/Desktop DM paths so they match Android behavior.
- commons ChatNewMessageState: actually attach the composed NIP-14 subject
to sent messages (the field was previously collected but dropped).
- commons ChatroomList: emit a `changes` SharedFlow on add/remove so list
UIs can refresh reactively; dedupe the User overloads onto the room ones.
- Desktop NewDmDialog: multi-recipient selection (chips + confirm button) so
a Desktop user can start a group, not only a 1:1.
- Desktop ChatPane/ChatroomHeader: show a group's NIP-14 subject in the
header and add a rename dialog that broadcasts a subject change to all
members.
- Desktop Main.kt DM ingest: route any ChatroomKeyable inner event into the
room (covers kind 14/15 and future variants) and store self-authored
NIP-37 drafts instead of dropping them.
- Desktop ChatroomListState: refresh reactively off ChatroomList.changes
(with a slower safety poll), track real per-room unread via a last-seen
mark, and hide rooms whose latest message isn't acceptable (mute/filter).
https://claude.ai/code/session_01VEukNczAYxNLBjLnqVEoZd
The commons commonTest source set is shared across all KMP targets,
including the iosArm64/iosSimulatorArm64 spike, but several tests still
reached for JVM-only APIs that don't resolve on Kotlin/Native:
- JUnit (`org.junit.*`, `junit.framework.TestCase`) → kotlin.test, with
message arguments moved from first (JUnit) to last (kotlin.test).
- `assertArrayEquals` → `assertContentEquals`.
- `@JvmStatic` on the `android.util.Log` test stub → removed (it only
affects JVM bytecode; companion calls work without it).
- `seg.javaClass.simpleName` → `seg::class.simpleName!!`.
- A test function name containing `()` (illegal on Native) → renamed.
- `String(CharArray, offset, count)` → `CharArray.concatToString`.
CliffDetectorTest exercises `computeStalledSpeakers`/`defaultCliffBackoffMs`,
which live in the jvmAndroid-only NestViewModel and are invisible to iOS,
so it moves to jvmTest alongside NestViewModelTest.
The in-stream loading marker spelled out "N relays" only for the fully-loaded
(done) chip; active frontiers showed a bare count ("Loading: ↓ 8"). Use the
relays plural for the count fallback on every state so a count chip always reads
as a sentence ("Loading: ↓ 8 relays"). 1–2 short host names still spell out.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
BackwardRelayPager exposed five independently-updated StateFlows (exhausted,
relayCount, stalledCount, reachedBack, relayProgress) that are all recomputed
together on every page settle. Co-located consumers therefore paid up to five
separate recompositions per settle and could observe a torn read (e.g. an
updated relayCount against a still-stale relayProgress).
Combine them into one atomic PagingStatus snapshot, emitted by a single
publish(), collected once. updateStatus()/recomputeExhausted() merge into that
publish() (exhausted computed inline). loadingMore stays separate: its falling
edge is debounced on its own timer in PerRelayLoadTracker, decoupled from the
status recompute, so folding it in would miss that delayed transition.
Threaded through the 3 history managers and the 3 feed consumers
(ChatroomListFeedView, ChatroomView, LoadingReplyNote): 12 collectors -> 4 at
the heaviest views.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The pre-existing desktopApp:UploadOrchestratorTest started failing
after the desktop image-compression feature landed:
- uploadCallsClientWithCorrectParameters (2x2 PNG, no quality set)
- uploadPassesAuthHeaderToClient (.txt)
- uploadPassesSameFileWhenNoStripExif (.txt)
- uploadComputesMetadata (.txt)
The orchestrator was unconditionally calling ImageReencoder.reencode,
which (a) reencoded PNGs to JPEG even when the caller did not opt into
compression and (b) threw UnsupportedFormat for any file the sniffer
could not classify (.txt, voice memos, video files, DM attachments —
the orchestrator is the upload path for everything, not just images).
Two changes restore the orchestrator's original "upload as-is"
behavior for callers that have not opted into compression:
1. UploadOrchestrator.upload's quality parameter is now nullable
(CompressionQuality? = null). Null means "do not reencode" —
matches the orchestrator's behavior before this feature, so the
Android, CLI, and any non-image upload path keeps working
unchanged. The desktop compose flow continues to pass a non-null
CompressionQuality so it still runs the reencoder.
2. ImageReencoder no longer throws UnsupportedFormat for
ImageFormat.Unknown — it returns PassThrough(NotAnImage) instead.
AVIF and HEIC still throw (those are recognized formats we
explicitly refuse). The new PassReason.NotAnImage is rendered in
the preview dialog as "Not an image · uploaded as-is" / "Metadata
preserved (non-image — no re-encode applies)".
My UploadOrchestratorTest.refusesAvifWithUnsupportedFormat is updated
to pass quality = MEDIUM explicitly so it still exercises the refuse
path under the new opt-in model.
Reported via runtime crash dialog on the user's first PNG upload:
Exception in thread "AWT-EventQueue-0":
java.lang.IllegalStateException: Can't overwrite cause with
javax.imageio.IIOException: Bogus input colorspace
at java.lang.Throwable.initCause(Throwable.java:464)
at CompressionException.<init>(CompressionException.kt:39)
Two real bugs:
1. CompressionException constructor double-set the cause.
Exception(message, cause) super already wires the Throwable's
cause slot; the init block then called initCause(cause) AGAIN
which throws IllegalStateException by spec ("Can't overwrite
cause"). The init block was added per a code-review note that
was wrong about how Kotlin's primary constructor forwards
cause. Removed the init block; relying on super does the
right thing. Regression: encodeFailedWrapsCauseWithoutCrashing.
2. JPEG writer rejected non-RGB BufferedImages with "Bogus input
colorspace". TYPE_INT_ARGB (typical PNG decode), TYPE_BYTE_GRAY,
TYPE_CUSTOM (CMYK JPEGs, indexed PNGs) all blow up the stock
JPEGImageWriter. encodeJpeg now flattens via toRgbCanvas — draws
onto a fresh TYPE_INT_RGB canvas with white background for any
transparent pixels. White matches what every major image viewer
does for transparent PNGs over a light surface.
Regression: reencodesPngWithAlphaToJpeg.
Both regressions are covered by new tests so the patterns cannot
silently come back. Reencoder test count: 13 -> 15.
The options row (Upload to / Quality / Post as) was wrapping
"Note" to two lines at the 600 dp dialog width — see the
screenshot the user surfaced during manual testing.
- Bumped the compose dialog from 600 dp to 780 dp and added
DialogProperties(usePlatformDefaultWidth = false) so the
explicit width is honored.
- Added maxLines=1 + softWrap=false to all three selector
TextButton labels (ServerSelector, QualitySelectorChip,
PostTypeSelector) so they can never wrap regardless of
future attachment count or label growth.
Also threads through a new preCompressed: File? param on
UploadOrchestrator.upload — landed early because the preview-
gate work needs it. When the upcoming CompressionPreviewDialog
hands off a pre-computed temp, the orchestrator skips reencode
+ stripExif and just uploads + cleans up.
Manual testing feedback from the UI:
- The Post-as Note/Picture FilterChip pair was overflowing the
options row and rendering "Picture" rotated 90°. Converted it
to the same Text + TextButton + DropdownMenu pattern as the
sibling Upload-to and Quality controls so the row stays
compact and visually consistent.
- QualitySelectorChip was likewise a FilterChip; rewritten as
Text + TextButton + DropdownMenu to match. Dropdown rows now
carry a two-line layout: bold preset label (e.g.
"Medium (640 px)") with a sub-line summary explaining the
tradeoff ("640 px · balanced size and quality").
- Restored the HIGH preset (640 px @ q=0.85, "visually lossless
on phones") between Medium and Desktop High. The four-preset
set is now Low / Medium / High / Desktop High — closer to the
Android-parity progression the brainstorm originally specced.
- CompressionQuality enum gains a `summary` field (one-line
description) plus a `chipLabel` convenience ("Medium (640 px)").
Settings panel and dropdown both consume `summary` so users
see what each preset actually does without trial-and-error.
ImageReencoderTest gains a HIGH-preset test and the monotonicity
test now asserts LOW < MEDIUM < HIGH for the same-dim subset.
Phase 7 of the desktop image compression plan.
Never silently downgrade — when ImageReencoder throws
CompressionException for one or more attachments (UnsupportedFormat,
InputTooLarge, EncodeFailed), the compose dialog now surfaces a
modal listing each failure and lets the user choose between Send
Original (uploads raw bytes, EXIF-stripped for JPEGs) or Cancel
post.
- UploadOrchestrator gains bypassReencode: Boolean = false. When
true, the orchestrator skips ImageReencoder and treats the
source as a PassThrough(BypassByUser), preserving the EXIF
strip + temp-cleanup semantics from the normal pass-through
path.
- ImageReencoder.PassReason gains BypassByUser.
- CompressionFailureDialog is built as Dialog { Surface } (not
AlertDialog) because:
* the body is a LazyColumn that grows with N failures —
AlertDialog's `text` slot has fixed-width constraints,
* LaunchedEffect / produceState don't fire inside
AlertDialog.text (see custom-feeds-alertdialog.md memory),
leaving room for future per-row actions (e.g. per-file
retry).
Each row shows: filename, "Could not compress: <reason>" in
error color, original byte count, and a privacy hint that
differs by source format ("EXIF will be stripped" for JPEG
bypass, "metadata may still be present" for PNG/HEIC bypass).
- ComposeNoteDialog send loop now wraps each upload in try/catch
(CompressionException), collects failures, and after the loop
awaits the user's choice via CompletableDeferred<FailureUserChoice>.
On SendOriginal: re-uploads each failure with
bypassReencode=true. On Cancel: aborts the post.
Phase 1 part C — the integration that makes the new compression
pipeline actually run on every upload.
- UploadOrchestrator.upload gains an optional quality parameter
(default CompressionQuality.DESKTOP_HIGH). The orchestrator now:
1. Runs ImageReencoder.reencode — branches on Reencoded vs
PassThrough.
2. For PassThrough + stripExif=true + JPEG source, runs
MediaCompressor.stripExif so animated/SVG passes still get
EXIF stripped where applicable.
3. Computes metadata on the bytes that will actually leave
the machine.
4. Eager-cleans up every intermediate in a finally{} block
wrapped in NonCancellable so user-cancelled uploads don't
leak temps.
Throws CompressionException for the fail-loud dialog path.
- MediaCompressor.stripExif drops deleteOnExit (long-running
desktop process was leaking stripped_*.jpg per
docs/temp-file-cleanup-analysis.md). Temp files now land under
AmethystTempDir with the amethyst_stripped_ prefix; caller owns
cleanup.
- AmethystTempDir lazy-initializer now triggers sweepOrphans() on
first access, recovering any amethyst_* files > 24h old left
behind by JVM crashes that skipped shutdown hooks.
- BlossomClient: open class + open upload methods so
UploadOrchestratorTest can substitute a FakeBlossomClient that
captures the uploaded file's bytes.
- jvmTest now pulls secp256k1.kmp.jni.jvm so end-to-end signer-
based tests (Blossom auth event signing) can run.
5 new orchestrator tests cover: JPEG re-encode path (uploaded file
is the AmethystTempDir temp, smaller than source, hash matches
captured bytes), animated GIF pass-through (uploaded == original),
AVIF refused with UnsupportedFormat (client.upload never called),
temp cleanup on success, and temp cleanup on upload failure.
Test totals across commons jvmTest upload package: 44 / 44 green
(4 smoke + 23 sniffer + 12 reencoder + 5 orchestrator).
Phase 1 part B. The core re-encode + downscale pipeline:
- AmethystTempDir resolves ~/.amethyst/tmp/ at mode 0700 with a
boot-time sweep of amethyst_* files > 24h. Defends against
/tmp tmpfs OOM on Linux VMs and against multi-user temp races
on shared systems. Overridable via -Damethyst.tmp.dir=.
- ImageReencoder.reencode(File, CompressionQuality) returns a
sealed ReencodeResult (Reencoded(file) | PassThrough(reason))
and throws CompressionException for fatal cases.
* Format sniffer first → pass-through for animated GIF /
animated WebP / SVG; refuse AVIF / HEIC with UnsupportedFormat.
* Pre-decode pixel guard: stream header dims via
ImageReader.getWidth(0)/getHeight(0), refuse > 50 MP before
any pixel buffer is allocated.
* Subsampled decode (floor stride) so a 4032×3024 source decodes
to ~2016×1512 in heap before Thumbnailator's final resize.
* Never upscale: Thumbnails.of(...).size() is only called when
the source actually exceeds the target box.
* CPU-bound work runs on Dispatchers.Default.limitedParallelism(1)
with ensureActive() between stages.
* Cleanup on cancellation/failure runs in NonCancellable.
12 unit tests cover preset routing, never-upscale, InputTooLarge,
AVIF/HEIC refused, animated-GIF pass-through, SVG pass-through,
JPEG SOI verification, and temp-file placement under
AmethystTempDir.
ICC profile preservation and the wide-gamut warning path land in
the next commit alongside UploadOrchestrator wiring.
Phase 1 part A of desktop image compression. Pure data types — no
external dependencies yet beyond Thumbnailator already on classpath.
- CompressionQuality enum (Low / Medium / Desktop High) with
JPEG quality values tuned for 2026 displays (0.65 / 0.75 / 0.90)
rather than the obsolete 2014-era Android values.
- ImageFormat sealed class covering the 9 formats the orchestrator
must distinguish (JPEG/PNG/BMP/TIFF re-encoded, animated GIF and
animated WebP byte-identical pass-through, SVG pass-through,
AVIF and HEIC refused for lack of a pure-Java decoder).
- ImageFormatSniffer with magic-byte detection + RFC 9649 VP8X
animation-flag check + GIF NETSCAPE2.0 application-extension
scan + ISO BMFF ftyp brand match for AVIF / HEIC variants.
- CompressionException sealed hierarchy (UnsupportedFormat /
InputTooLarge / EncodeFailed) with initCause for chain
preservation through logging.
23 sniffer unit tests cover each format including JPEG-via-file,
empty input, missing file, and both WebP animation paths.
Phase 0 of the desktop image compression plan
(docs/plans/2026-06-08-feat-desktop-image-compression-plan.md).
- commons jvmMain gains net.coobird:thumbnailator:0.4.21 (pure-Java,
MIT) — to be consumed by the new ImageReencoder in Phase 1.
- amy CLI now sets -Djava.awt.headless=true via three paths so any
transitive ImageIO/AWT touch never spawns a GUI thread:
* applicationDefaultJvmArgs in cli/build.gradle.kts (covers the
installDist startup scripts and any future jpackage launcher),
* the amyImage custom Unix launcher in cli/build.gradle.kts,
* System.setProperty as the first line of cli Main.kt — belt-
and-braces for invocations that bypass the launcher scripts.
- commons:jvmTest forces -Djava.awt.headless=true for the same
reason during test runs.
Smoke tests (CompressionSmokeTest.kt) document Thumbnailator's
upscale-by-default behavior — ImageReencoder must gate the resize
itself in Phase 1.
Some sources (e.g. the math-academy posts) over-escape their LaTeX, so
the content carries `\\ldots` / `\\cdots` instead of `\ldots` / `\cdots`.
JLaTeXMath reads the `\\` as a forced TeX line break — splitting every
inline equation across two lines — and renders the trailing command name
as the literal letters "ldots"/"cdots", so the ellipsis symbol is lost.
Collapse doubled backslashes (`\\cmd` -> `\cmd`) before rendering inline
math; display math is left untouched since `\\` can be a genuine line
break there. No-op when the content isn't over-escaped.
The "where the live tail ends and paged history begins" boundary (one week)
was copy-pasted across the gift-wrap + NIP-04 live tails, the backward history
pager floor, and the prune's recent/old split — easy to drift out of lockstep
(overlap = double-load, gap = missed messages).
- Add DmHistoryTuning: one place for liveTailSeconds + recentKeepCount, read by
AccountGiftWrapsEoseManager, both NIP-04 SubAssemblers, BackwardRelayPager,
and Chatroom.pruneMessagesToTheLatestOnly. Drops the duplicated
LIVE_TAIL_SECONDS / DEFAULT_LIVE_TAIL_SECONDS constants.
- Log the prune-time window realignment under DMPagination, per scope
([giftwrap] / [rooms.nip04] / [convo.nip04]): relay count + newest pruned
timestamp, so the rewind is observable in logcat.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Pins that a MathSegment covers only its $-delimited span (not the
paragraph), that space-separated hashtags, URLs and images next to math
stay independently detected, that `$x$.` keeps its trailing period while
a following hashtag remains its own segment, and that currency `$5`
doesn't pair with a later equation. Also documents the glued (no-space)
edge cases.
Memory pruning drops DM messages out of the cache but left the per-relay
paging cursors untouched, so a relay still claimed to have delivered the
dropped band (reachedUntil deep, or done) and the demand-driven loader
never re-requested it — a silent hole until app restart.
- Prune NIP-17 too: pruneMessagesToTheLatestOnly now reaps both NIP-04
(PrivateDmEvent) and NIP-17 (WrappedEvent rumors) on one merged top-N
cut, so a conversation is cut at a single time point (no NIP-04-without
-NIP-17 holes). NIP-17 is the actual memory-pressure driver.
- HostStub carries the host's createdAt, so a decrypted rumor self-
describes its outer gift-wrap time (the time the cursor pages by; the
rumor's own time is the message time, not the wrap time).
- RelayLoadingCursors.rewindTo() pulls a relay's reached cursor up past
the pruned band, clears done, and un-arms it (demand-driven re-fetch);
advance() now resumes from the rewound reached point instead of the
floor.
- LocalCache.pruneOldMessages accumulates the newest pruned created_at
per relay (outer-wrap time for gift wraps, event time for NIP-04),
filtered below each cursor's floor, then rewinds giftWrapHistory +
rooms-list nip04History (account-wide) and the per-conversation
nip04History.
The gift-wrap window is account-global, so pruning one room rewinds the
shared sweep; the interference is bounded (already-held wraps short-
circuit in consumeRegularEvent, re-fetch is demand-gated).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Vitor PamplonaandGitHub
davotoula
nrobi144
Claude