Review prep for the DM pagination branch:
- Delete commons TimeWindowPagination + its test: the early since-based
time-window approach, referenced only by its own test and fully
superseded by UntilLimitPager (until+limit, gap-proof). 212 lines a
reviewer would otherwise study for nothing.
- Bring the design doc up to the final architecture: NIP-04 per-relay
filter scoping, per-relay independent paging (no rounds) + in-stream
markers for the convo, the round model still used by rooms/gift-wrap,
the WindowLoadTracker backstops and tracksReqSends gating, the
loadingMore-starts-false fix, and the DMPagination diagnostics map.
Marks the obsolete time-slice section as superseded.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
loadingMore was wired straight to windowLoad.loading, which starts true
(the tracker assumes a load is in flight from construction). On the
first conversation open — before any paging window has run — that true
wedged the scroll-driven loader: its gate is '!loading', so loadMore
never fired, and even if it had, the 'if (!windowLoad.loading.value)
startLoading' guard would have skipped startLoading (value was the
construction-time true), leaving no watchdog to ever settle it. Result:
permanent spinner, 0 relays. Earlier opens only worked because a prior
conversation had left the shared tracker at false.
Expose a _loadingMore that starts false and is mirrored from the window
by the done collector, and track windowActive ourselves so the first
loadMore actually starts the window (and a re-entrant loadMore mid-page
doesn't reset it and forget finished relays).
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Kind 34551 (CommunityRulesEvent) was missing from EventFactory.create, so
signing a community-rules template produced a generic Event. Returning it
as CommunityRulesEvent in Account.sendCommunityRules threw a
ClassCastException when publishing community rules.
Register the kind in the factory and add a regression test.
- Mark all *_search_keywords translatable="false" (English concept/protocol
index; stops Crowdin translating protocol terms and breaking locale search) [#1]
- Collapse ~23 symbol+nav rows via a local symEntry() helper [#3]
- Reword keywordsRes KDoc to match the actual word-prefix tokenization [#6]
- Add search keywords to the Legal rows (privacy_policy, child_safety) [#7]
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add keyword blobs so settings resolve by concept/protocol name, not just
title — e.g. "blossom" -> Media Servers, "audio rooms" -> Nests Servers,
"negentropy" -> Event Sync, "nsec" -> Backup Keys. NIP numbers omitted by
preference.
- Collapse filterSettings' two identical lookup lambdas into one stringLookup
- Rebuild via SettingsCategory.copy() so new fields aren't silently dropped
- Memoize buildSettingsCatalog with remember(hasPrivateKey, nav, uriHandler);
onResetMarmot reads isResettingMarmot via rememberUpdatedState to avoid a
stale-capture, eliminating ~60 allocations per keystroke
feat(settings): add search box that filters settings rows by title + keywords
refactor(settings): expose legal section as legalSettingsCategory factory
feat(settings): add search placeholder, empty-state, and keyword strings
test(settings): cover filterSettings (blank/title/keyword/category/empty/danger)
refactor(settings): match category title in search; use data classes
Two related polish fixes on the collapsed sidebar:
1. The hover/active highlight on each nav item used to span the full
sidebar width (minus 8dp outer padding), producing ~12dp of empty
highlight either side of the 24dp icon. Now the highlight clips to
a 40dp square centered on the icon (24dp icon + 8dp padding on each
side), so the ripple sits tight against the glyph.
2. When the sidebar is collapsed, the label was already supplied as
`contentDescription` for screen readers but had no visual
affordance. Added a `TooltipArea` that surfaces the label on hover
(Surface + inverseSurface tonal style, matching the existing
TorStatusIndicator tooltip pattern), so mouse users can also see
what each icon means without expanding the sidebar.
Applied to both `SidebarNavItem` and `SidebarFeedItem` since both
suffer the same issue. Expanded behaviour is unchanged.
Two bugs that together caused HomeFeed to always open on Following:
1. FeedScreen was reading feedRepo.pinnedFeeds.value as the source of
truth for the first pinned feed. That's a stateIn-derived flow with
initial value persistentListOf(); the underlying _feeds StateFlow
IS loaded synchronously by FeedDefinitionRepository on construction,
but the derived pinnedFeeds doesn't reflect it until the first flow
emission propagates — which is too late for `remember` to see.
Fixed by reading feedRepo.feeds.value directly and filtering /
sorting by pinOrder ourselves.
2. DeckColumnContainer was passing initialFeedMode = FeedMode.FOLLOWING
when rendering DeckColumnType.HomeFeed, which overrode FeedScreen's
first-pinned logic entirely. Removed the hardcode so the deck's
home column inherits FeedScreen's default.
With both fixed, a user who has only Global pinned now opens to Global
on launch instead of Following.
If the user has pinned only Global (or only a custom feed), the app
should open to that on launch instead of showing Following just
because DesktopPreferences.feedMode happened to be saved as
Following. The "pinned feeds" list is the user's stated ordering;
the first item should drive the initial tab.
Resolution order (most specific wins):
1. explicit customFeedSource/customFeedId from the caller
2. explicit initialFeedMode from the caller
3. first pinned feed in feedRepo.pinnedFeeds (NEW)
4. DesktopPreferences.feedMode (last-saved, previous default)
For a pinned Filter feed, this also seeds activeFeedId and
activeFeedSource so the feed mounts in CUSTOM mode with the right
source.
Real root cause of the "stale feed on launch" perception bug: when
fresh events prepend to the desktop home feed, Compose's stable-key
diff (`items(loadedState.list, key = { it.idHex })`) preserves the
visual anchor on whatever item was already visible. The user's
previously-visible top item — once at index 0 — silently shifts to
index N as N new items are inserted above the viewport. From the
user's perspective the feed looks frozen on stale items even though
the underlying state HAS updated; switching screens unmounts
FeedScreen, recreates lazyListState at index 0, and on remount paints
from the now-current top.
Android already handles this with StickToTopOnPrepend
(amethyst/.../WatchScrollToTop.kt:133-152), but the helper lived in
the Android module and Desktop had no equivalent.
Changes:
- New commons/.../ui/feeds/StickToTopOnPrepend.kt with the same
observer + snapshotFlow trick, ported to use plain `collectAsState`
(replacing the Android-only `collectAsStateWithLifecycle` — the
effect's lifecycle is already bound to composition via
LaunchedEffect). Provides the same overloads:
* StickToTopOnPrepend(LazyListState, firstItemKey)
* StickToTopOnPrepend(LazyGridState, firstItemKey)
* StickToTopOnPrepend(FeedContentState, LazyListState)
* StickToTopOnPrepend(FeedContentState, LazyGridState)
- FeedScreen wires StickToTopOnPrepend(viewModel.feedState,
homeFeedLazyListState) at the same scope as the hoisted lazy list
state and the NewPostsChip.
Mutually exclusive with the NewPostsChip: the chip's visibility
predicate fires when isAtTop is false, the auto-snap fires when
isAtTop is true. Together they cover both cases:
* user at top → events arrive → auto-snap shows them
* user scrolled down → events arrive → chip announces them
The Android version in amethyst/.../WatchScrollToTop.kt is left in
place to avoid a wider refactor; it can be reduced to a thin delegate
in a follow-up.
Both loading splashes (the Tor-connect gate and the account-loading
screen between Tor active and LoginScreen) now show the Amethyst
icon tinted to the theme primary, anchored below the status text.
Layout pattern (status-forward, both splashes):
spinner → status text → Amethyst logo (96.dp, primary tint)
Brief research summary backing the choice:
- Apple HIG argues against splash branding, but its model assumes
near-instant launch — not applicable here where the Tor gate
can block for seconds.
- Material Design 2's branded-launch-screen pattern endorses
logo + brand color while a placeholder UI loads.
- The status-forward order keeps the dynamic info (what we're
waiting on) leading and the brand as the anchor below — the
right call when the wait is non-trivial.
Audit follow-up on the per-relay paging work:
- Restore DmRelayLog in the convo history loadMore (every other nip04 /
giftwrap assembler logs it) and add per-relay milestone logs: which
relay reached the bottom, which stalled and why, plus a one-line
done/still-trying breakdown when the window settles — the snapshot to
reach for when a chat doesn't load.
- Extract markStalled() (dedupes onClosed/onCannotConnect, logs once per
relay) and relaysFor() (dedupes the active-convo relay lookup).
- relayCount now counts the relays still being paged (done ones drop out)
instead of staying frozen at the total.
- Drop the unused loadEverything().
- Fix WindowLoadTracker docs/log that claimed it 'gives up' on silent
relays: it only stops waiting and reports them; the owner decides (the
convo keeps them open and retries). Fix a dangling KDoc link in
RelayReachMarker.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Visualises the per-relay-independent engine: each relay gets a thin
marker in the message stream at the depth (createdAt) it has paged down
to, sitting just below the oldest message it has loaded. As a relay
pages older its reached-back cursor drops and the marker slides down; a
relay that races ahead leaves its marker deep while slower relays' trail
higher and converge as they catch up — ✓ done (empty-EOSE), … stalled
(auth CLOSE / unreachable, still trying), ↓ reaching. Hidden once the
conversation is fully converged (every relay done or stalled).
Adds an optional markersInGap slot to the shared chat feed view
(no-op for public-chat callers) invoked per message gap with its
createdAt bounds; ChatroomView renders the markers from the
relayProgress flow.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Replaces the lock-step round model (every relay advanced one page per
global round, gated by the slowest) with per-relay continuous paging:
each relay continues to its next page the instant it EOSEs, off its own
cursor. The subscription layer diffs per relay, so re-issuing only
re-REQs the relay whose cursor moved — others' in-flight REQs are
untouched. Fast relays race to the bottom in back-to-back pages while
slow / auth-walled relays catch up at their own pace; none are
abandoned (this reverts the give-up behaviour — slow relays keep their
subscription open and keep trying), so every relay converges on the same
window.
A relay is done on an empty page; one that won't answer (auth CLOSE,
unreachable, silent) is marked stalled but keeps trying. loadingMore
clears once every relay is done or stalled. Exposes per-relay
RelayPagingProgress (reached-back / done / stalled) for the upcoming
in-stream progress markers.
Splits the convo widen loop so each protocol pages on its own loader
state — NIP-04's continuous loading no longer starves gift-wrap paging.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
The silence + connect-grace backstops are only meaningful when the owner
feeds onReqSent, which only the convo NIP-04 path does. But connectStalled
keyed off the ABSENCE of a recorded REQ, so for giftwrap/rooms (which never
call onReqSent) every relay looked connect-stalled after connectGrace —
the window completed at ~15s before its REQs had even gone out during a
slow connect storm ('giftwrap.live load done: settled/silent' 45s before
the REQ), prematurely declaring an empty round done and tripping the
no-progress guard, so giftwraps stopped loading.
Gate both REQ-aware backstops behind a tracksReqSends flag that only the
convo manager sets; everyone else keeps the plain settle / idle / cap
behavior. Adds a regression test that a non-tracking tracker keeps
blocking a never-heard-from relay until it actually settles.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
After the connect-grace fix the convo history no longer hangs, but it
could linger forever showing 'N relays' with no progress bar and
exhausted=false: a correspondent's auth-walled relay (ditto, which only
ever CLOSEs 'all authors must be authenticated') never reaches the
pager's done/given-up state, and the no-progress guard merely *skipped*
re-issuing the round — stopping the loop without ever reflecting that
we're finished.
When the guard trips (same active relays, zero events two rounds
running) give up on those relays and recompute exhaustion, so the
conversation reports as fully loaded and the relay count clears. My own
reachable relays empty-EOSE to 'done' and never reach this branch, so
only genuinely stuck correspondent relays are dropped.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
The convo history spinner could sit for the full 5-minute absoluteCap
('load done: cap' in the logs). A correspondent relay (ditto) dropped
after its auth-required CLOSEs and got stuck reconnecting on a flaky
network, so its round-2 REQ was never delivered. It therefore reached
neither a terminal signal (no CLOSE without a REQ) nor the silence
backstop (which measures from REQ-delivery), and blocked the round until
the cap.
Add a connect-grace backstop: a relay that has been expected past
connectGrace (15s) without even receiving its REQ — i.e. stuck
connecting — stops blocking the round. Unlike the silence backstop it
does NOT give the relay up (it may be a genuinely slow connect), so the
owner keeps it and retries it next round; only relays that accepted a
REQ and then went silent are abandoned.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
The convo NIP-04 history spinner stayed up for minutes against 9 relays
even when the conversation was fully loaded. Auth-walled relays (ditto,
nostr.wine, …) accept the REQ (success=true) and then send nothing — no
event, no EOSE, no CLOSED. WindowLoadTracker only completed once every
relay reached a terminal signal, an idle gate that required hearing from
*all* relays, or the 5-minute cap; a silent relay armed none of those,
so only the cap freed the spinner. The pager likewise kept the silent
relay 'active' every round, so the count never dropped and exhaustion
never completed.
Add a silence backstop keyed off onSubscriptionStarted (REQ delivered,
post-connect — so a slow connect isn't mistaken for a dead relay): a
relay that received its REQ but stays silent past silenceTimeout (10s)
no longer blocks completion and is reported via onAbandoned, which the
convo assembler uses to giveUp() the relay in its pager so it leaves the
active set and lets exhaustion finish. finish() applies the give-up
before flipping loading, so the round collector recomputes exhaustion
after the silent relays are dropped.
Tests cover the pager give-up/exhaustion and the tracker's silence +
connection-gap behavior.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Nip04DmRelays held two flat relay sets and every filter named the whole
conversation group, so a relay that belongs to only one counterpart was
still asked about all of them (e.g. {authors:[bob,charlie]} sent to a
relay that is only charlie's). Restructure it into per-relay key maps so
each relay sees exactly the keys it owns:
fromMe: my outbox -> {authors:[me], #p:[whole group]}
each counterpart inbox -> {authors:[me], #p:[keys reading there]}
toMe: my inbox -> {authors:[whole group], #p:[me]}
each counterpart outbox-> {authors:[keys publishing there], #p:[me]}
Relays shared across roles union their key sets, so my own relays still
carry the full group while a counterpart's relay only ever names that
counterpart.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
A conversation's NIP-04 relay set folded the correspondent's inbox
(read) relays into the from-me filter set, so we sent {authors:[me]} to
relays that belong to the other party (e.g. ditto). Those relays have no
reason to hold my authored messages and auth-walled ones reject the
filter outright ("all authors must be authenticated"), stalling the
load. Scope filters to relay owners: my outbox carries my messages; the
correspondent's outbox (plus my own inbox as a legacy safety net)
carries theirs. Drops groupInbox from the from-me set.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Adds DmRelayLog, a diagnostic that prints — per DM subscription — the
account whose relays are being used and breaks the relay set down by the
source list each relay comes from (NIP-65 inbox/outbox, DM-relay-list,
private-storage outbox, local relays). Wired into all six DM assemblers
(NIP-17 live/history, NIP-04 rooms live/history, NIP-04 convo live/history).
The existing REQ lines now also print the resolved relay URLs (split into
fromMe/outbox and toMe/inbox for the NIP-04 paths), so an unexpected relay
— e.g. a write-only NIP-65 relay that only the NIP-04 home+dm path queries —
can be traced back to the list it leaks in from.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
With "exhausted only when every relay empty-EOSEs", a relay that demands auth we
can't satisfy — e.g. relay.ditto.pub answering "auth-required: all authors must
be authenticated" for a correspondent's pubkey we can't authenticate as — CLOSEs
every round, never finishes, and the conversation loads forever.
UntilLimitPager now tracks a per-relay CLOSED streak; after GIVE_UP_AFTER_CLOSES
(3) consecutive CLOSEDs with no answer in between (so the pool's auth handshake +
a retry have already failed), the relay is marked "given up" and excluded from
activeRelays. It is NOT counted as done (it didn't empty-EOSE — we just can't
read it), but it no longer blocks exhaustion. The streak resets on any event or
EOSE, so a relay whose auth succeeds is never abandoned. onClosed wires into the
pager and re-checks exhaustion when a relay tips into given-up.
NIP-17 only queries the user's own DM relays (auth = self), so it rarely hits
this; it's the NIP-04 conversation fan-out to the correspondent's relays that
trips author-auth-required relays.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Exhaustion was decided by "the round returned zero events," but a round can
return zero because relays CLOSED (auth) or never answered — not because they
reached the end. So "All caught up" appeared before slow/auth relays had
actually finished.
A relay is finished only when it returns an empty page followed by EOSE — the
pager already records exactly that in its per-relay `done` flag (CLOSED /
cannot-connect deliberately don't set it). So the chat is now exhausted only
when every relay is done (activeRelays is empty), per the rule "all relays must
return that EOSE." A post-auth empty EOSE that lands after the round already
settled on the earlier CLOSED now flips exhausted immediately too
(markExhaustedIfAllDone in onEose), not just at the next round boundary.
Because a chat is no longer "done" while a relay keeps CLOSING, the conversation
auto-fill (no stall-gate) would otherwise re-issue identical rounds and hammer
that relay. A no-progress guard skips re-issuing a round whose relay set and
zero-event result are unchanged; the pool re-auths on the open subscription and
its EOSE clears the guard and finishes the relay. All the new single-valued
state resets on account/conversation switch alongside the existing display flows.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Replaces the "pubkey/listId" string key with a small ConvoKey data class holding
the account pubkey and the ChatroomKey. ChatroomKey is a data class over the
participant set, so it's a collision-free key — unlike listId, which is its
32-bit hashCode as a string and can collide. Including the account keeps the two
accounts' views of the same correspondent on separate cursors (the manager is a
singleton shared across logged-in accounts).
Also lighter on allocation than the string it replaces: the per-relay-event hot
path captures the key once per subscription (no per-event construction), and the
remaining call sites build one small object instead of concatenating + hashing a
string.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
The DM managers live in a single shared coordinator (Amethyst.instance.sources),
so every logged-in account uses the same instances. The per-account paging state
(pager cursors, started set, accounts map) was keyed by pubkey and fine, but the
single display flows — exhausted, relayCount, reachedBack, autoFillRoomMark —
were not. Switching from an account that had exhausted its history to another
left exhausted=true, so the second account's auto-fill was gated shut and its
chats never paged in.
Each history manager now tracks the active account/conversation and repoints its
display flows on switch: exhausted is restored per-account (kept in a small
exhaustedByUser/exhaustedByList map so an already-finished account shows "all
caught up" rather than re-paging), and the cosmetic flows + stall mark reset.
Paging cursors stay in the per-account pager, so progress is preserved.
Also scopes the conversation history pager key by account pubkey: a ChatroomKey
(hence listId) is identical for the same correspondent across accounts, so two
logged-in users viewing the same person would otherwise share one cursor.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Replaces the bare spinner + "Load entire history" link with a status card that
tells the user what the app is actually reaching for: per protocol, it shows
"Older <encrypted|legacy> messages" with a subtitle of "<NIP-17|NIP-04> · N
relays · back to <month>" while it pages. When that protocol runs dry the card
doesn't just vanish — it crossfades to "All caught up · Reached the start of
your <…> messages", holds for a beat, then collapses away.
The history managers now surface the live status the card needs: relayCount
(relays the current page is asking) and reachedBack (oldest point paged to, from
the deepest per-relay cursor), added to all three history managers and computed
via UntilLimitPager.deepestUntil. The "load entire history" action is dropped —
scroll-driven paging already walks to exhaustion, so the link was redundant.
Each protocol's card sits at its own oldest-loaded boundary (rooms list) or both
stack at the conversation's oldest end, so the two protocols' loading is shown
independently at their real depths. New strings use a <plurals> for the relay
count.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
The rooms list had a single auto-fill trigger and a single loading boundary,
both pinned to the oldest private room of EITHER protocol. When NIP-04 history
ran far deeper than NIP-17 (e.g. NIP-04 back to 2023, NIP-17 shallow), that
oldest room was a 2023 NIP-04 row at the very bottom, so gift-wrap loadMore only
fired when the user scrolled all the way down to it — NIP-17 never paged on the
way, and the loading indicator was only visible at the bottom.
Split the trigger and the boundary per protocol. Each protocol now widens on its
OWN oldest-loaded room (gated only on its own loader and its own room-count
stall-gate) and shows its OWN loading indicator at its own depth, so NIP-17 and
NIP-04 page independently as the user scrolls — matching their very different
histories. WidenPrivateWindowWhen is generalized to a per-protocol WidenHistoryWhen
called once per protocol (and once per protocol for the empty-feed hunt). Each
history manager carries its own auto-fill stall mark (autoFillRoomMark).
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Fixes the perceptual "stale feed on launch" bug: on cold launch the
desktop feed paints with whatever local cache had (up to 7 days old)
before relays catch up. The live updateFeedWith() path already prepends
fresh events silently, but users had no signal that fresh content
arrived unless they were already at the top of the feed (auto-snap via
StickToTopOnPrepend).
This adds a Twitter/Mastodon-style floating pill chip that slides down
from above the search header when fresh events have prepended AND the
user is scrolled below position 0. Tapping it smooth-scrolls to top
and slides the chip back up off-screen. Scrolling to top manually
also dismisses it.
Implementation:
- NewPostsChip + rememberNewPostsChipState in commons/commonMain so any
future feed surface (incl. Android, iOS) can adopt it. Desktop wires
it today; Android continues with the existing auto-stick + bottom-nav
dot pattern.
- Visibility predicate is pure-function and unit-tested (5 cases).
- Predicate mirrors the inverse of StickToTopOnPrepend's "at top" check
so the two systems are mutually exclusive — auto-snap when at top,
chip when not.
- Chip placement: floating Alignment.TopCenter inside FeedScreen's outer
Box, offset by the animated headerSpacerHeight (60.dp normal,
300.dp when search is expanded) so it tracks the header card.
- Hoisted lazyListState + headerSpacerHeight one level so the chip can
share scroll state with the LazyColumn. Existing viewport-aware
metadata loading is unchanged (same lazyListState reference).
- Animation: slideInVertically(tween(280, FastOutSlowInEasing)) + fadeIn
for enter; slideOutVertically(tween(220, FastOutLinearInEasing)) +
fadeOut for exit. Initial/target offset of -fullHeight-16 guarantees
the chip is fully off-screen above its rest position.
- Per-column scope by construction: each FeedScreen instance has its
own chip state (deck mode shows one chip per column).
- Resets cleanly on feed mode switch (Following ↔ Global ↔ Custom)
because rememberNewPostsChipState is keyed on FeedContentState,
which is recreated when viewModel = remember(feedMode, activeFeedId)
recomposes.
Plan: docs/plans/2026-06-02-feat-new-posts-chip-desktop-feed-plan.md
5 issues from davotoula's review on PR #3124:
- #3 (protocol): inline reply emitted a minimal e/p tag set instead of
NIP-10. Extract `commons/actions/ReplyActions.replyTo` wrapping
`TextNoteEvent.build(replyingTo=)` (which already encodes root marker,
reply marker, parent root-e-tag carry) + carry parent's p-tag chain via
`notify(...)`. Replies to deep-thread notes now thread correctly in
Damus/Primal/Coracle. Covered by `ReplyActionsTest`.
- #4 (architecture): reaction/follow/reply each inlined
`localCache.consume + relayManager.broadcastToAll` in 5 sites with
inconsistent ordering. Extract `desktopApp/cache/dispatch(...)` —
canonical local-first order — and route all 5 sites through it.
- #1 (UX): related-content section scanned the cache once via
`DisposableEffect(noteId)` and never refreshed. Switch to `produceState`
collecting `DesktopLocalCache.eventStream.newEventBundles`; re-scan only
when an arriving bundle contains a candidate (matching hashtag or
author). `LargeCache.notes` is a ConcurrentSkipListMap (weakly consistent
iterator) so the scan stays safe on the composition coroutine.
- #2 (UX): `DeckColumnContainer` re-requested focus on every
`currentOverlay` change, stealing focus from sibling columns whenever
any column mutated overlay state. Drop to `LaunchedEffect(Unit)` and
wrap the column in `key(column.id)` in `DeckLayout` so the one-shot
effect survives column reordering.
- #5 (consistency): zap totals bypassed the shared `ZapFormatter`. Wire
`RelatedContentRow`, `CommentItem`, and `NoteActions` to
`commons/util/ZapFormatter.{showAmount,toZapAmount}`; delete
`formatZapAmount` and `formatSats` desktop-local helpers.
`WalletColumnScreen.formatSats` intentionally kept — locale-aware full
precision for wallet balance is by design.
Plan: docs/plans/2026-06-02-fix-desktop-feed-review-findings-plan.md
Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
1. LocalRelayStore: use batchInsert() with per-row savepoints instead of
manual transaction — UNIQUE constraint violations skip that row instead
of failing the whole batch
2. Robohash empty hex: guard blank input in CachedRobohash.get() with a
fallback all-zeros hex key instead of passing empty string to assembler
3. GiftWrapEvent decrypt: downgrade from WARN to DEBUG — expected when
gift wraps from local relay cache aren't addressed to current user
(subscription filter is correct, but hydration doesn't filter by p-tag)
4. Relay URL %20: decode percent-encoded spaces before rejection check in
RelayUrlNormalizer.fix() — wss://relay.example.com/%20 now normalizes
to wss://relay.example.com/ instead of being rejected
5. NIP19 Parser: downgrade from ERROR/WARN to DEBUG — malformed bech32
from relay content is expected in the wild, catch+log is correct
6. VLC macOS: add --avcodec-hw=none (disables VideoToolbox that causes
CVPN chroma failures) and --reset-plugins-cache (rebuilds stale cache
on startup instead of logging hundreds of stale-cache errors)
Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>
The time-slice history bounded re-downloads but couldn't tell "this relay is
empty" from "this is a gap" — an empty time slice can sit above older messages,
so the only stop was the 10-year maxLookback, and a wide late slice could pull a
20k-event firehose in one request.
History now pages backward by until+limit, per relay (UntilLimitPager). Each
round asks every not-yet-empty relay for up to 10000 events older than its own
cursor, no since, so gaps are skipped: an empty page + EOSE is a gap-proof
"nothing older on this relay" signal. A relay returning fewer than the limit is
treated as its own cap, not exhaustion — only an empty page ends it. A relay
answering CLOSED isn't "empty" (it may answer after the auth handshake), so the
global exhausted flag flips only when a whole round advances no relay at all,
which also stops the loop on a relay that keeps CLOSing. limit caps per-request
volume too.
Both NIP-04 history managers now paginate themselves (per relay, scoped) instead
of following the gift-wrap slice; loadEverything pages to the end by auto-issuing
the next round until exhausted. The live tail and the rooms-list stall-gate are
unchanged. Filter builders gained an optional limit; the conversation NIP-04
helper exposes its outbox relay set + a per-relay until builder.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
The conversation auto-fill only fired when the thread overflowed the screen, so
a one-message room sat at its load-more boundary without ever advancing — you
were at the start of the chat but it wouldn't reach for older messages. That
overflow guard was added back when each widen re-downloaded the whole window
(to stop a short thread auto-walking the gift-wrap firehose); now that history
loads in bounded, non-re-downloading slices that reason is gone.
Drop the overflow requirement: load the next slice whenever the oldest end is in
view, including a thread too short to scroll. A one-message room now walks
history back to its real beginning (or until the window is exhausted), one
bounded slice at a time, gated on both loaders being idle.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Every widen re-requested the whole DM window (the filters carried `since` only,
no `until`), so a relay re-streamed the entire history from the new floor — a few
pixels of scroll walked the window to the 10-year backstop, re-downloading
exponentially more each step (589 → 1486 → 2609 events in one session). This
splits each DM protocol into two responsibilities:
- Live tail (existing managers, now fixed): a one-week floor with no `until`,
always open to the future. Never widens, so new messages keep arriving.
- History slices (new managers): load the past in bounded `since`+`until`
one-shot slices. Widening fetches only the new band `[newFloor, prevFloor]`;
consecutive slices are disjoint so advancing the filter never re-streams an
earlier slice — they live in the cache. The NIP-17 2-day wrapper-timestamp
margin is applied to the slice `since`, overlapping adjacent slices so a
randomized outer timestamp can't open a gap. NIP-04 (exact timestamps) needs
no margin.
New: AccountGiftWrapsHistoryEoseManager owns the geometric window and the
bounded slices; ChatroomListNip04HistorySubAssembler / ChatroomNip04History-
SubAssembler follow its slice bounds so both protocols page to the same depth.
The live managers (AccountGiftWrapsEoseManager and the NIP-04 followers) are
reduced to the fixed one-week tail.
Also adds the rooms-list stall-gate: the auto-fill remembers the private-room
count at the last widen (on the history manager, so it survives reopening the
screen) and stops widening once a step brings in no new private room — widening
pulls older messages, not rooms, so a few busy correspondents would otherwise
flood events without ever filling the list. "Fill until full OR nothing new
found", instead of walking to the 10-year backstop.
Design: amethyst/plans/2026-06-01-dm-live-tail-and-history-slices.md
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
The wire-level diagnostics logger only recognized gift-wrap kinds (1059/21059),
so kind:4 NIP-04 REQs never produced a `REQ -> wss://…` line and their relays'
connect/CLOSED/NOTICE lines were filtered out — even though the kind:4 REQs are
issued (the `[rooms.nip04] REQ` manager logs show them going out). Broaden the
match to the whole DM path (1059/21059/4) so the wire trail covers both
protocols, and rename the gift-wrap-specific identifiers to dm-path.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
The load summary's "before floor" count compared incoming gift wraps against the
un-margined window floor (window.since), but filterGiftWrapsToPubkey actually
asks relays for `since = window.since - 2 days` to catch wraps whose randomized
outer timestamp dips below the real message time. So the deliberate 2-day margin
band showed up as "before floor" (a boot reported "6 before floor" that were all
legitimate margin-band wraps), conflating the intended margin with a relay that
ignores `since`. Compare against window.since - twoDays() so only a relay that
under-shoots the floor we actually requested is flagged.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
A boot trace reported `[giftwrap] load summary: 1 event(s)` for a 7-day window
that actually holds ~100, because the DM relays connect over a ~35s spread: one
fast relay delivered a single event at +7s, the next 3s were quiet only because
the other four relays were still mid-connect, and the idle heuristic mistook
that gap for "done". The rest streamed in afterwards, past the load boundary.
The clean "all relays answered" completion was also unreachable: relays that
answer `CLOSED auth-required` (and unreachable relays) never produced an EOSE,
and WindowLoadTracker ignored onClosed/onCannotConnect entirely — so the only
completion path was the too-eager idle timer firing in a connection gap.
Completion is now per-relay terminal-state based. A relay is "settled" once it
sends a terminal signal — EOSE, CLOSED, or cannot-connect — and the load is done
when every targeted relay has settled. This is fast when relays are fast
(everyone EOSEs in a couple seconds) and correctly patient when they are not
(waits for the slowest relay), and it cannot trip in a connection-stagger gap.
The idle timer is kept only as a backstop for a relay that streams without ever
EOSE'ing, gated behind "every relay has been heard from" so it too can't fire in
a gap; the absolute cap still bounds a relay that connects then hangs forever.
WindowLoadTracker.trackingListener now wires onClosed and onCannotConnect into
the tracker, and a live event no longer settles a relay (its preceding EOSE
does); forward (newEose) semantics are unchanged.
https://claude.ai/code/session_01B1fmmmX8JjQWH3amMLdvcW
Claude
Vitor PamplonaandGitHub
David KasparandGitHub
Crowdin Bot
nrobi144