| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| ... | |
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A draft is written as plain text but doesn't arrive as plain text: a URL
ending in .jpg becomes a picture, a youtu.be link becomes a player, a
nostr: reference becomes the note it points at, an npub becomes a name.
The only way to find out how yours would land was to send it.
The eye beside the paperclip — on the composer and on a reply — opens a
preview. It runs the draft through noteBody, the very renderer the feed
uses, so this isn't a lookalike of the real thing: it IS the real thing,
in a dashed frame with a PREVIEW label so it can never be mistaken for
the post itself. It follows the draft live while it's open (and only
while it's open — a render per keystroke is not free).
Attaching a picture opens it by itself. Showing you what you attached is
the answer to the question attaching it raises, and it costs no tap.
tools/compose-preview-test.js covers both the composer and a reply: the
picture renders rather than sitting there as a URL, the YouTube link
becomes a player, the preview tracks typing, and it puts itself away.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Tapping a face opened an empty page that grew a banner, then a bio, then
a column of posts as relays answered — measured as the browser measures
it, a cumulative layout shift of 0.412, where 0.25 is already "poor".
None of those fetches needed to happen on the tap:
The bio and banner were free all along. The batched kind-0 fetch that
names everyone in the feed downloads their WHOLE profile and then throws
away everything but the name, picture, nip05 and lud16. It keeps `about`
and `banner` now, and the page reads them until the full kind 0 lands —
after which `full` is the truth, including when it says there is no bio.
The posts were in memory. The feed you just tapped through holds up to
two hundred notes and an open thread holds its replies; the ones by this
author paint immediately while the relay fetch gets their real page.
And the rest is a head start: an avatar (or the name beside it) begins
its profile's fetches on pointerdown rather than click. That is a hundred
odd milliseconds free, and unlike the prefetch-every-row this replaces —
sixty requests for twenty rows, which is what starved the profile batch —
it only ever asks about the person actually being tapped.
Measured after: layout shift 0.000, bio in the first frame, posts in the
first frame. tools/profile-open-test.js asserts exactly that, and fails
on the old code with 0.412 and a blank page, so it cannot quietly rot.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Measured on a Pixel 9: Amethyst gives a feed avatar 55dp, we gave it 30 —
the same circle the inbox and chat rows use, where a face is a label on a
line. In a feed the face is the point. Feed rows now get 44px; everywhere
else is untouched.
The catch, and the reason this isn't a one-line change: the cached
thumbnail was 96px square, sized for a 30px circle at 3x. A 44px circle
no longer fits inside it, and the fallback is to layer the ORIGINAL over
the thumbnail — which for a feed of twenty rows means twenty full-size
downloads, one of which in this wallet's own cache was a 5MB JPEG. That
is exactly the slow-avatar problem we already fixed once.
So the thumbnail grows with the circle: 144px, which covers 44 at 3x with
room to spare. Faces cached at the old size are remade once (thumbPx says
which size a thumbnail was made for), and the old one keeps painting
meanwhile, so nobody watches a grey circle while it happens. The per-face
cap goes 9000 → 16000 chars and the shared budget 180k → 280k, both
scaled to the new size: the busiest picture on hand measures 12827 chars
at 144px, and at the old cap it would have been rejected and left
fetching its original forever.
avatar-thumb-test gains the guard that would have caught all of this:
the thumbnail must cover the biggest circle in the app at 3x. It failed
at 128px (132 needed) — which is why THUMB_PX is 144 and not 128. Its
redirect assertion was also order-dependent — the seed writes the profile
into every wallet's cache but only the painted one makes a thumbnail, so
it asked the first row and got a coin toss. It takes the best copy now.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A feed of posts was set in .small — 13px, the size used for captions and
row subtitles. On a phone it read like footnotes. Measured against
Amethyst on the same Pixel 9, side by side: its note body renders a line
pitch of 62 device pixels, ours 50.
Post bodies get their own class at 16px, which is Amethyst's own default
for a note body, and chat and DM bubbles move 15 → 16 to match; the names
and inbox previews above them keep pace. New line pitch measured on the
phone: 60 against Amethyst's 62. A quoted note inside a post stays a step
smaller, since it's context rather than the post you're reading.
Worth recording for whoever reads this next: this phone has the Android
font scale at 1.15. Amethyst uses sp and honours that automatically; CSS
px in a TWA does not, and there's no way for a page to read that setting.
So Amethyst will still render about 15% larger than us on this device no
matter what we pick. Matching its DEFAULT is the target that makes sense
for everyone else.
chat-layout-test had rotted (onboarding copy, .header-msgs now matching
the settings button, a screenshot path from another machine's scratch
dir) — repaired, and it's what proves the composer still fits at the
larger size, on all three viewports.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A screenful of eighty rows sent eighty kind-0 REQs in the same breath.
Relays cap how many subscriptions a connection may have open, most were
refused, the five-second timeout expired, and the empty answer was then
believed for ten minutes — which is a real account dressed as a punk with
a shortened npub for the rest of the session. Measured against this
wallet's own feed, 131 of the 133 people on screen had a kind 0 sitting
on relays we were already connected to. It was never reach; it was us.
So profiles go out in batches: a 250ms beat collects a paint's worth of
rows into one request of up to 200 authors, and each batch paints as soon
as it lands rather than waiting behind the outbox fallback chasing
stragglers. A miss now backs off instead of sticking — fifteen seconds,
then half a minute, up to the ten minutes that's right for someone who
genuinely has no profile.
And noteRow no longer calls prefetchProfilePage. That fetches an author's
whole page — kind 0, relay list, last thirty notes — which for twenty rows
is sixty requests, and is what starved the batch in the first place. A row
is not a profile tap; the page still loads when the profile is opened.
profile-feed-test had rotted in three places (onboarding copy, and
.header-msgs now matching the settings button first) — repaired, not
rewritten.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Compared against a wider net over the same six hours and the same follow
list — 862 people, read off this wallet:
before: 165 notes from 38 people
after: 327 notes from 86 people, and nothing the wider net still finds
Four things were losing posts, in order of how much:
THE AUTHOR CAP. feedAuthors() stopped at 500. On an 862-follow list that
silently dropped 362 people, and 164 of the 166 posts the wider net found
were theirs. Gone: it takes everyone now, up to a couple of thousand, with
each REQ carrying at most 400 authors so no single filter is absurd.
THE DEFAULTS. relay.damus.io is the second-biggest relay this follow list
writes to — 249 of them — and it wasn't in our set. 232 follows publish no
relay list at all and lean entirely on those defaults, so they were being
asked for on two relays instead of three.
THE SOCKET CAP. Ten relays covered all but 66 follows; twenty-five covered
all but 26. It's 24 now — Amethyst runs near a hundred, and this is the same
idea kept to a number a phone can hold.
MALFORMED RELAY LISTS. One list in these follows crams three URLs into a
single r tag, space-separated. It passed our "starts with wss://" check,
reached the pool, threw Invalid URL — and every author covered by that relay
quietly returned nothing. Each URL has to parse now.
And a quote-post is no longer mistaken for a reply. NIP-10 marks a reply's e
tags 'root' or 'reply'; a quote's are 'mention', and NIP-18 quotes carry a q
tag. We were hiding every note with an e tag, which included the quote-posts
that now render as cards.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A nostr:note1/nevent1 in someone's post IS a post, and making the reader tap
a link to see it — losing their place in the feed — was the wrong shape. It
renders inside the post now: whose it is, when, and what it said, in a card
of its own so the two voices don't run together. Tapping it opens the note
properly.
Deliberately not a noteRow: a quote is context, not a second post to act on,
so it carries no reply, boost or zap of its own. The action row stops at the
post that did the quoting.
Fetched once per id and remembered, so a feed quoting the same note ten times
asks for it once. One level deep only — a quote of a quote of a quote is a
rabbit hole, and the inner one stays a link you can follow.
tools/quote-card-test.js quotes a real note off the relays and checks it
arrives with its author, its text, and none of the outer post's buttons.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
All three shapes it comes in — youtube.com/watch?v=, youtu.be/, and a Shorts
link — become the video's still with a play button on it and a corner link
out to YouTube. A ?t= keeps its start time. Anything that merely looks like
one (example.com/watch?v=…) stays a link.
The player itself arrives on the first tap, not on render. A feed of ten
videos would otherwise load ten YouTube players, each one telling Google what
you're scrolling past before you've decided to watch anything — and on a
phone, ten players is the scroll. One tap and it plays in place, from
youtube-nocookie.com.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A push sent to ntfy.sh reached a killed app on a real phone:
17:02:21.259 ntfy.sh → the phone's ntfy → broadcast to io.coinos.webview
17:02:21.308 app process started (it had been killed)
17:02:21.458 wake: job started
17:02:22.926 wake: finishing (page said so) — 1.7s, and "You were paid"
One thing that took two goes: `adb shell am force-stop` marks a package
STOPPED, and Android delivers no broadcasts to a stopped package unless the
sender sets FLAG_INCLUDE_STOPPED_PACKAGES — which `am broadcast` does and a
real distributor rightly doesn't. The first attempt looked like a failure and
was a test artefact; `am kill`, which is what Android itself does to a
backgrounded app, delivers fine.
And the WebView build can now say something useful about push. It was
inheriting the browser's line — "your browser has no push service, try
Firefox" — which is nonsense inside a WebView that IS the app. It asks the
host what's actually there and says: install a push app like ntfy (nothing
installed), or wait a moment and reopen (a distributor that hasn't answered
yet). Payments and messages still arrive whenever coinos is open, which is
the part that matters most to someone reading an apology.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The reply box gets the same paperclip the post composer has: one uploader
now, handing the URL to whichever draft asked for it, and the reply's own
imeta tag describing what it carries.
And the UnifiedPush build is graphene.apk — what people will actually call
it, and short enough to read down a phone line.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Web push on Android is Firebase's, so a phone without Google Play Services —
GrapheneOS, CalyxOS, any de-Googled Android — cannot be woken by a web app at
all. UnifiedPush is the way that doesn't involve Google: the user runs a
distributor (ntfy and friends), apps ask it for an endpoint, and whoever
wants to reach the device POSTs to that endpoint.
THE WRAPPER speaks the protocol directly — four broadcasts each way, so it
pulls in no library; an APK this small has no business growing a Kotlin
runtime for eight intents. It asks the distributor on every launch (the call
is idempotent, and it's how a distributor installed later gets noticed),
raises a notification when a message arrives, and acknowledges it so the
distributor stops resending.
THE ENDPOINT reaches the web app on the launch URL as ?up=, the same road
payment intents already travel — because the distributor talks to the app and
only the web app knows which pubkeys are worth watching. It's registered with
the notifier in place of a browser subscription, for notifications and for
NWC alike, and needs neither permission nor push service.
THE NOTIFIER sends to either kind: one deliver() that POSTs the payload to a
UnifiedPush endpoint, or hands it to web-push as before. Same payloads, same
404/410 pruning. The endpoint is the only secret there, so it's never logged
in full.
Verified: a POST to a live ntfy endpoint delivers the payload byte for byte,
so the app parses exactly what the notifier sent; the APK builds, carries the
receiver, and is signed with the certificate assetlinks already pins (v1.6).
tools/unifiedpush-test.js drives a browser with no push service and checks
the endpoint is taken off the URL, kept, and registered.
Honest about the limit: a UnifiedPush message wakes the ANDROID app, not
Chrome's service worker, and the wallet's keys live in the browser's storage
where nothing here can reach them. So a zap request arrives as a notification
you tap, and the wallet answers when it opens — one tap instead of none.
Automatic answering while closed still needs a browser with a push service.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
"Registration failed - push service error" is Chrome's own wording, and it
tells nobody anything. The cause is that Chromium routes web push through
Firebase: on GrapheneOS — or any Android without Google Play Services —
there is no push service to register with, so the app cannot be woken while
it's closed, and no amount of tapping the toggle will change that.
Both places that ask for push (the chat's offer and the NWC "answer while
closed" toggle) now recognise that specific failure and say: this browser has
no push service; on Android that's usually a phone without Play Services;
Firefox brings its own and works there; and zaps are still answered whenever
coinos is open. The last part matters most — the feature isn't broken, it's
narrower than they thought.
The failure is remembered so the app stops retrying it on every launch, and
an explicit tap on the toggle always clears that and tries again, because a
phone that changed browsers deserves another go.
tools/push-unavailable-test.js runs a browser whose subscribe() fails exactly
as Chrome's does and reads back what the user is told.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Reply and zap have come down from the header to join boost, quote and the
heart, spread across the post's full width so each is a comfortable target
instead of five buttons huddled in a corner. Only the ellipsis stays up top.
Proper icons at a readable size, all the same weight: a speech bubble for
reply (a return arrow is what a keyboard does, not what a person does), the
repeat arrows for boost, quote marks, a heart, a bolt. Counts sit beside the
icon they count.
And the heart asks which feeling: tapping it opens the same six emoji the
chat sheet offers, and the one you pick is the one the post then shows —
NIP-25 takes any content, so there was never a reason to assume a heart.
Tapping your own reaction still takes it back.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Under every post now: a heart with its count, a boost with its count, and a
quote. Behind an ellipsis beside reply and zap: quote, boost, copy the note's
reference, mute the author, or block them. Likes and boosts come back in the
SAME relay request the zap tallies already make — one round trip answers all
three — and each action paints before any relay does, the way the zap chip
learned to.
A like is kind 7 and a boost kind 6, published to the author's own write
relays as well as ours, so the person being answered actually sees it. A
quote opens the composer with a nostr:nevent reference already in the text
and tags the post on publish, so other clients render the embed. Taking a
like back is the only withdrawal nostr has — a deletion request naming the
reaction.
Mute is the NIP-51 list (kind 10000): their posts leave the feed at once and
the list is published, so it follows you to other clients. Block is the same
list plus an unfollow, which is why it says so.
And a bug the mute found, which the follow list had too: the optimistic paint
was being read back as fact. Both toggles paint the answer immediately and
then publish onto the newest list the relays will give them — but when the
relays had nothing newer, "the newest list" WAS the optimistic paint, so the
code saw the follow already present and published a list without it. Every
follow was a no-op that looked like it worked. The base is now the relays'
copy only when it's genuinely newer, and otherwise what we held before the
paint.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The attach control was a .btn-sm holding a 16px icon, wedged between two
full-height buttons — a sliver to aim at on a phone. It's a 48px square the
height of the row now, with a 20px clip in it.
And tapping an image opens it full screen: the image centred on a dark
ground, tap anywhere (or Escape) to close. Posts crop images to 320px so a
feed stays a feed, which makes a photograph a thumbnail — this is where you
actually look at one.
Two details in the viewer. The page underneath stops scrolling while it's
open, and the scrollbar gutter the page keeps reserved stands down with it:
fixed positioning is sized to the viewport WITHOUT that gutter, so an
"inset: 0" backdrop leaves a pale stripe down the side of an otherwise
full-screen picture.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
It was rendering everything it held — a hundred-odd notes to show the few you
read. It now paints twenty, and reaching the bottom shows twenty more; only
when the window catches up with what's cached does it go back to the relays
for older posts. Posts arriving live at the top widen the window rather than
pushing off the one you'd scrolled to.
And the posts are the page now, not a card sitting inside one: they step out
through the chat page's padding to the container's own edge, leaving one
continuous list with hairlines between the posts instead of two sets of card
walls.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Our set is three relays — relay.coinos.io, nos.lol, relay.primal.net — plus
whatever the user's own NIP-65 names. Asking those for everyone's posts only
finds the people who happen to use them; anyone else looks like they stopped
posting.
So the feed now reads NIP-65: relay lists are fetched in batches of a hundred,
kept for a week, and the authors covered greedily — the relay carrying the
most people not yet covered, repeated up to ten sockets. A feed of hundreds
is a handful of connections that between them reach everyone, rather than one
each.
Our relays are still asked about everyone, not just the people with no list.
Measured across eight busy accounts, each source holds posts the other misses:
reading the author's own relays surfaced 7 to 40 notes ours had never seen
(jb55 +40, verbiricha +25, pablo +23), and ours held plenty theirs didn't —
fiatjaf's own three relays answered 42 where ours answered 67. The feed wants
the union, so the outbox adds reach rather than replacing it. On the test set
that took the feed from 83 posts to 118.
Profile pages get it for free: relaysOf now reads the batch's cache first, so
opening someone you follow costs no extra round trip.
The test records every socket the page opens and checks the plan by what it
actually dialled.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
FOLLOWING. A profile page gains Follow / Following beside DM and Pay. The
contact list is the kind-3 event, kept locally so the button is right in the
first frame and refreshed from relays behind it. Publishing one is delicate:
a contact list is a shared document — other clients keep relay hints in its
content and petnames in its tags — so a follow re-fetches the newest list
first and adds or removes exactly one p tag, preserving everything else. The
button flips immediately and flips back if the publish fails. The service
worker's friend filter reads the same list, so following someone also means
their DMs stop being treated as a stranger's.
THE FEED. Posts by the people you follow, newest first, at the top of the
messages home. Replies are left out: a reply belongs to its thread, and a
timeline of half-conversations reads like eavesdropping. The last screenful
is kept locally so it opens with posts in it rather than a spinner, new posts
arrive over a live subscription while you're looking at it, and older ones
stream in as you reach the bottom — the same shape as the profile pages. A
wallet that follows nobody is told so and offered the people search.
MEDIA. The composer (now shared by the feed and the profile page) takes a
picture or a video: it uploads when you pick it and appends the URL to the
text, which is how every nostr client reads media, with a NIP-92 imeta tag
published beside it for the ones that would rather be told than guess. Posts
already render images and video inline, so they show up where they land.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| | |
|
| | |
|
| | |
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Pinguimdarelva's avatar still flashed, and the reason was one I'd have kept
guessing at: from the live browser, `new Image()` with crossOrigin set FAILS
for that picture, while `fetch(url, {mode:'cors', cache:'reload'})` for the
very same URL returns 200. The host allows us perfectly well.
The picture sits behind a 302, and the avatar is painted as an ordinary
background-image — so the bytes are already in the HTTP cache under the
redirect's target as a no-CORS entry, and the crossOrigin <img> reuses that
entry and fails the CORS check. The image loads, the CORS copy "errors", and
the face gets written off.
So the thumbnailer fetches the bytes itself and draws from a blob: no
crossOrigin element, no cache collision, and an untainted canvas without
depending on the element's CORS bookkeeping. With no taint left to detect,
failures no longer split into hard and soft — each one just backs off, from
six hours up to a fortnight, so a host that truly won't have us is asked
twice a month instead of every boot.
The test now serves a picture behind a redirect and loads it plainly before
thumbnailing it, which is the shape that broke; it fails on the old code.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Pinguimdarelva's picture is 52KB on a CDN that answers in 700ms one minute
and times out the next — I measured both from this machine. The thumbnail
attempt hit one of the bad minutes, and `thumbFail` was written as if the
host had refused us: never retried, so that avatar went back to fetching the
original on every boot.
Only a tainted canvas is a real refusal (no CORS header, nothing to be done),
and that is remembered for good. Everything else — a load error, or a host
that won't answer inside twenty seconds — is stamped with the time and tried
again in six hours. The timeout also frees the slot: three thumbnails run at
a time, and a hanging image used to hold one of them indefinitely.
And zap tallies are remembered between sessions. They were asked of the
relays on every boot, which takes seconds, so a room full of zapped messages
painted bare and filled in afterwards. The totals (never the receipts) are
kept locally and painted straight away; the moment the relays answer for a
message, their word replaces it — so a tally that changed, or a receipt that
was deleted, still settles correctly. Opening a DM thread doesn't forget what
the chat had learned, and the store is capped at 400.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The avatar delay left on the messages page wasn't the image at all. A person
with no cached profile got an empty circle until a relay answered — and the
one that stayed blank longest (npub1k6f89cg, in the live cache) is never
cached at all: persistProfile deliberately refuses to write a row with no
name and no picture, so every boot re-asked the relays and every boot sat
blank for seconds. When it finally arrived it was a punk, which is why it
looked like slow art rather than a slow lookup.
The punk is derived from the pubkey alone, so it can be drawn in the first
frame. It's already the right answer for everyone with no picture, and for a
coinos user who kept the default it IS their published picture — same
formula, same file. A real photo replaces it when it lands. A name lookup in
flight still gets the quiet circle: a picture is expected there, and punk art
must not flash in front of it.
Also: the punk rewrite now recognises the bare relative path a few older
profiles carry, not just the absolute URL.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The art is 240x240 (20KB each, 1.5MB for the set) because that's what a punk
published as someone's nostr picture should be — but the circles the app
draws are 18 to 64px, so a new wallet with nothing cached was fetching eight
times the pixels it could show, one request per face.
punks-sm carries a 128px copy of each: 4KB instead of 20KB, indistinguishable
at 64px. Small circles and the wizard's picker tiles load those; only the 64px
profile avatar and the value the wizard publishes stay full-size.
The bigger win is that a punk picture is our own art wherever it turns up.
Everyone who keeps the default publishes v3.coinos.io/punks/N.webp as their
nostr picture, so those faces now paint straight off disk at the size being
drawn — no request, no thumbnail to make, and no broken circle when the
legacy coinos.io copies 502 (which they do for a third of the set).
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A profile picture is whatever its owner uploaded, and that is very often the
full-size original. Measured against the faces this wallet had cached: one
avatar is a 4,988,208-byte JPEG, another 3,901,966 — megabytes fetched and
decoded to paint a 30px circle. That is the second of blank circles on the
messages page after a refresh, and it repeats every time the upstream cache
header expires.
So the first time we see a picture we downscale it once (96px, webp) and keep
the thumbnail beside the profile in localStorage, which is read synchronously
at boot. Small avatars then paint from it alone and never ask the network for
the original again; the 64px profile avatar layers them, so the thumbnail
shows instantly while the full-size original loads over the top.
Reading the pixels needs CORS — most avatar hosts allow it (6 of the 9 here),
and the ones that don't are remembered as failures rather than retried every
boot. A refresh of the profile carries the thumbnail across when the picture
hasn't changed. Thumbnails live on a 180k budget, oldest face first, and Data
Saver opts out of making them at all.
tools/avatar-thumb-test.js watches a wallet meet a large cross-origin picture
and checks the next boot fetches no picture at all.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The optimistic amount is supposed to stand down when the real receipt is
counted, and it did — but only by deleting itself. An ark zap publishes its
own receipt, which comes back over the live subscription BEFORE the send
call resolves, so the flow's success report arrived second and re-created
the amount on top of the receipt already in the total. 21 paid, 42 shown.
The pending entry now has a state (flying / paid / void) instead of a
deletion: our receipt voids it, and a settle arriving afterwards can't
resurrect it. A receipt older than the tap — the backlog the first paint
fetches for a message zapped before — doesn't void anything, so zapping the
same message twice still counts both.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
A zap took a few seconds to pay and its public receipt trailed that by
seconds more, so the tally chip appeared long after the tap — the tap looked
ignored, and the only way to tell it had worked was the toast.
Now the chip appears on the tap carrying your amount, breathing while the
payment is in flight; it goes solid the moment the payment lands, and steps
aside when the real receipt arrives (which is also how it avoids counting
your zap twice). The ark and Lightning one-tap flows report their outcome
back — a failure, or a fallback into the form, takes the chip straight off
again rather than leaving it pulsing. An unreported zap gives up after 45s.
A zap sent from the form reports too, so the tally shows it right away
instead of waiting on a relay. Reduced-motion gets a dimmed chip, no pulse.
Co-Authored-By: Claude Opus 5 (1M context) <[email protected]>
Claude-Session: https://claude.ai/code/session_019mrt228Upn9nNCwGixzSLx
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Chat and DM message sheets grow a ⚡ Zap item (not on your own messages),
riding the same one-tap flow as a post zap — Ark first, Lightning fallback,
the amount remembered from the first time. A chat message is a rumor with an
id like any event, so the zap e-tags that id; the public receipt reveals
nothing but the id.
Every post row, chat bubble and DM shows how much it's been zapped as a
little amber bolt + sats chip (1.2k / 21k / 1.5M), fed by the public receipts
that e-tag it: NIP-57 kind 9735 (amount from the bolt11, the zap request's
amount as fallback) and coinos' own Ark receipt, kind 9737 (network-checked).
Summed per id, deduped by receipt; underlined when one of ours is in it.
Ids are asked once each (batched per paint, 150 per REQ), and ONE live
subscription follows the ~200 ids most recently on screen so a zap landing
while you watch — yours included, rechecked at 4/12/30s — shows up without
a reload. The tally chip is itself a zap button on others' messages.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Claude-Session: https://claude.ai/code/session_01URmfo8VimCNkw6Y7JndYrv
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The Accounts screen is about a seed's Spending/Savings faces — its rows say
nothing about WHO each pair belongs to, so a "Sign into another account"
button there was the wrong door. It's gone from there. On your own profile,
"Sign into another account" now opens a switcher: every identity already
signed in on this device as an avatar + name row (one per seed, wearing the
nostr key its header avatar shows — the login key when there is one, else
the seed's own; remembered on the account and in the durable directory by
xpub, so vault-restored accounts have a face too), tap to switch with nothing
logged out, plus "Sign into a new account" for the additive sign-in doors.
With no other identity on the device the button goes straight to those doors.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Claude-Session: https://claude.ai/code/session_01URmfo8VimCNkw6Y7JndYrv
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The additive sign-in door lived only at the foot of the Accounts screen; the
profile page — where people actually go looking for a way out of the current
identity — offered just Log out. Now it carries the same button, via a shared
signInAnother action on the feature ctx (clears the profile/chat state so the
unlock screen's Back lands on the wallet, and the new identity doesn't open
onto the old identity's profile).
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Claude-Session: https://claude.ai/code/session_01URmfo8VimCNkw6Y7JndYrv
|
| |
|
|
|
|
|
|
|
|
|
| |
The new-post composer held its text in a transient ui field, so the reload
that reconnects a signer (or any reload) lost the post. It now rides the
same persisted draft store under a 'post' key: the composer reopens with
the text after a reload, posting clears it, Cancel closes the composer and
keeps the text for next time.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Nw1Dijik1tLVKRM1WkU7xE
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
field), not a toast.
A pasted-nsec login is never stored, so a reload leaves the link without
a signer. The post, reply and profile-edit paths checked for an identity
by THROWING the "signer isn't connected" message, which the composer then
caught and toasted — bypassing noIdToast, whose whole job is to open the
reconnect screen where the nsec (or extension / bunker) can be re-entered.
Those paths now go through requireIdentity(): it opens the screen (or
toasts, when there's no link to reconnect) and throws a silent error the
callers don't repeat.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Nw1Dijik1tLVKRM1WkU7xE
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
relay-only sync records.
A remote-signer (bunker) wallet re-opened the relay's last ~400 gift wraps
on EVERY boot — the seen-set lived in memory only — and each wrap costs two
NIP-46 round trips with a schnorr sign+verify apiece on the client. On the
Pixel that was ~4s of a 15s pegged boot. Opened (and given-up) wrap ids are
now kept in the messages feature state, bounded to 1500 id prefixes, and
loaded before the inbox subscription.
The outbox also replaces a record whose route changed even when its digest
matches: the 874KB relay-only ark record that could never publish would
otherwise have stayed in front of its Blossom successor forever.
Verified live: the wallet's 633KB ark domain uploaded to blossom.coinos.io
and its pointer was acknowledged by relay.coinos.io on the next boot.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Nw1Dijik1tLVKRM1WkU7xE
|
| |
|
|
|
|
|
|
|
| |
A wallet has two nostr keys — the seed-derived key and, when signed in, the login identity (extension/bunker/passkey/Google). Sending already used one identity, but RECEIVING/advertising used both: startDMs subscribed kind-1059 gift wraps on '#p' for both keys and fetched both keys' inbox relays, and ensureDmRelayList always published a kind-10050 inbox for the seed key even under a login. Result: a login wallet advertised the seed key as a SECOND DM inbox, so the welcomer greeted both keys (duplicate welcome) and a person's DMs split across two identities — one showed up as if from "another user".
Collapse to a single DM identity: new dmSubKeys() returns just the login pubkey when signed in, else the seed key, and startDMs subscribes/fetches only that; ensureDmRelayList no longer advertises the seed key when a login identity is active. isMe() and dmDecryptors() stay tolerant of both keys so already-received messages still thread and align. Not a privacy fix — nothing ever leaked to a third party (both keys were the user's own); this removes the confusing second self-identity going forward. (Existing cached DMs to the seed key on a given device don't retro-clean; a chat re-sync/forget clears them.)
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
Reactions were a sibling of the bubble (a separate row under it), so they hung outside the message. Moved the reaction pills inside .chat-bubble, tucked under the text, across all three renderers: community channel messages, DMs, and the read-only public /chat page. The pill background and the "your reaction" highlight border now derive from the bubble's own text colour (color-mix on currentColor), so they read correctly on both the plain bubble and the accent "me" bubble. Tapping a pill stops propagation so it toggles the reaction without also opening the message action sheet.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
notifications went silent. registerPush reused whatever getSubscription() returned and only ever subscribed fresh when there was none; a subscription made against an OLD VAPID key can never receive a push (the notifier's send fails with a permanent 403, which was never pruned, and the client kept re-registering the same dead endpoint). Now the client fetches the current VAPID key first and, when an existing subscription's applicationServerKey doesn't match, unsubscribes and re-subscribes against the current key — so a device self-heals on next open. And the notifier prunes 403 endpoints alongside 404/410. Deployed nwcpush on cs
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
stops opening the staging PWA. The same seed derives the same nostr pubkey on mainnet and staging, so both PWAs registered push under one pubkey and the notifier — matching only by pubkey — fanned every mainnet payment out to the staging subscription too; tapping it opened staging.coinos.io. Now the client stamps its network on the push registration (pushWatch net: getNetwork()), the registrar declares its own network on the /notify call (CFG.forwarder.network), and nwcpush delivers a payment only to registrations whose net matches — with a rollout-safe fallback: a device that hasn't re-registered yet (net=null) still gets the push, and a caller that omits net still reaches everyone, so nobody goes dark mid-deploy. Deployed names + nwcpush on cs
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
community rooms and DM threads (now sharing a chat-page class) drop their card walls and run edge to edge under 520px — and the type steps up a notch: bubbles 14→15px, names 13→14px, timestamps 11→12px. Plus: a fixed-price LNURL (minSendable == maxSendable — bolt-card top-ups, paywalls, static charges) no longer prompts for an amount — resolve() fills it and locks the field via the same fixedSat machinery fixed-amount BOLT 12 offers already use (disabled input, adjusted hints, confirm bypass)
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
client tag — ['client', 'coinos'], name-only since no kind-31990 handler coordinate is published yet — so other clients show "via coinos". Deliberately absent from everything encrypted (DMs, community wraps): private traffic needs no billboard
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
chunk, so every boot painted bare heads until it arrived — but the core bundle can't carry the hat art. Instead, wrapAvatar now leaves a STAMP per head in localStorage (art + percent geometry, size-independent, capped at 40, cleared when a head goes bare so removals can't ghost), and the core avatar factory replays stamps until hatsReady answers — from then on a null from wrapAvatar genuinely means bare. Verified headless: a stamped hat is on the header avatar 15ms after DOM-ready, and once the chunk lands with a bare-head answer the stamp is cleared with no double-wrap
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
sits INLINE, directly under the note it answers, instead of at the bottom of the thread where nobody scrolled to find it — slotted after the focused note (the root by default), appended at the end only until the focused note loads, with a refocus handshake so the cursor survives the box relocating when the thread finishes loading. (2) Posting is OPTIMISTIC: a provisional dimmed copy tops the feed before any signing or relay round-trip (a remote signer alone can take seconds), swapped for the signed event on success and withdrawn on failure — the composer closes immediately and reopens with the text intact if publishing fails; pending rows can't be opened/replied/zapped until real. (3) The full header nav (search, chat, settings, lock, avatar) now shows on the profile, thread, and search screens whenever a wallet is open — losing the search button on a profile made finding the next person a trek back home; the public no-wallet surface keeps the plain brand. All verified headless: box renders mid-card under the focused note with focus landing after load, a post appears 250ms after tapping Post (pre-publish) and survives settle, and the search button works from a profile
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
the post's thread focused on that note and puts the cursor straight in the reply box (inside an already-open thread it just re-targets the reply to that message). Verified headless on /adam: 51 rows all carry the button, tapping one lands with the thread reply input focused
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
pulls the next page of older kind-1s — until = oldest seen, one fetch in flight, a short answer marks the end; the listener registers at feature construction so the public no-wallet /username pages scroll too), VIDEO rendering in noteBody (mp4/webm/mov/m4v as <video controls preload=metadata playsinline> beside the existing inline images), and FULL-BLEED on phones — the notes card drops its walls and runs edge to edge under 520px; the inner 440px scrollbox is gone everywhere, the page itself scrolls. Verified headless at phone width on /adam: feed spans 0px-to-edge both sides, images render inline, and scrolling grew the list 51 → 82 posts
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
Delete — a gift-wrapped DM can't be retracted from the peer's relays). DM reactions ride the kind-7-in-a-gift-wrap shape 0xchat uses (makeDMReaction in dm.js), sent to the peer and echoed to our own inbox so other devices fold them in; inbound kind-7 rumors land in dmReacts keyed by the reacted message's id — no thread lookup, so a reaction arriving before its message still sticks. Replies e-tag the quoted kind-14 (makeDMRumor grew extraTags) and render the same accent-bar quote as community chat; the composer gets the same replying-to bar. Chips are joinable and your own reaction is highlighted/replaceable, all optimistic with rollback. Round-trip verified: reply and reaction rumors wrap/unwrap with tags intact
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
profile page immediately, not after Save: while the editor is open the DRAFT is the truth for the banner, the avatar, and the display name shown up top — clearing a field previews the removal too. The editor's text fields render per keystroke so pasted URLs track live (the morph never rewrites a focused field, so it can't fight the typing); uploads already rendered after filling the field. Verified headless: typing picture/cover URLs into the edit form updates the page-top avatar and banner backgrounds before any save
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
— it was hidden on your own page), bio, and latest posts, exactly like anyone else's, with the editor behind an explicit Edit profile button (Save or the new Cancel returns to the view) and a New post button that unfolds a composer — publishPost signs a plain kind-1 with the login/wallet identity, publishes to the identity's own relays, and prepends it to the notes list so it shows immediately. The header avatar's tooltip says "Your profile" instead of "Edit profile" to match. Compose/edit state clears on open/back so drafts never leak across profiles. Verified headless end-to-end: onboard → avatar → view mode (buttons + npub, no form) → Edit opens form → Cancel → New post opens composer
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
bubble opens a centered sheet: a quick row of the six reaction emojis (your current one highlighted; picking fires immediately) above Reply / Copy / Delete (delete on own messages only) — replacing the hover-only 🙂 picker, which was invisible on the phones that need it most. Links, images and the hover × keep their own clicks, and a desktop text-selection drag never pops the sheet over the copy. REPLIES are new: Reply arms a composer bar showing who/what you're answering (× to cancel, cleared on channel/room switch), the sent kind-9 e-tags the quoted message, and rows carrying an e-tag render the quoted context above their text — author + one-line snippet, silent when the original is deleted or not held. The public /chat page renders the same quotes. Verified against the built stylesheet: 300px sheet, six even emoji buttons, quote ellipsizes, circular cancel button
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
DM bubbles now run through noteBody (URLs → anchors, image URLs inline, npub mentions clickable, all content kept as text nodes — nothing from a relay reaches innerHTML); the public /chat page gets its own linkedBody (same splitter, nostr refs stay dim stubs since there are no profile screens there). Reactions: a 🙂 opener on every bubble (rides the edge like the delete ×, steps left on own rows) unfolds a six-emoji picker; tapping an existing chip joins that vote, your own reaction is highlighted, and picking a different emoji replaces it (kind-7 rumor on the channel stream, one per author, fold keeps the latest — inbound rendering already existed, this adds the sending half). Optimistic like sendMessage: the chip lands instantly and is rolled back to whatever stood before if signing/wrapping fails
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
| |
|
|
|
|
|
| |
that shows the live chat history to anyone, logged in or not, with a "Sign in to chat" button ("Open coinos" when the device already has an account). No server-side decoder needed: the community's join material was already public by construction (baked into the client bundle — every coinos user is meant to be a member), so the page derives the stream keys and decrypts wraps client-side exactly like the app's Chat tab, worker-offloaded with the same sliced fallback. Renders messages + edits + deletes + reactions and honors the folded control state's ban list; presence/typing/guestbook deliberately omitted. Profiles come from ONE batched kind-0 query per burst (a REQ per author tripped relay concurrency limits). Plumbing: the join material moved to src/community.js (shared with messages.js so they can never drift), /chat is excluded from the single-segment profile deep-link (which was eating the path and rewriting the URL before the route ran), and app renders stand down in public-chat mode so a feature-init render can't paint onboarding over the page. Verified headless against the live relays: 59 messages decrypt, ban-fold applies, names resolve
Co-Authored-By: Claude Fable 5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01MGzU12QrAA4G6BYGzGVdXz
|
