| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
| |
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01QjVvE4YAcK3AGwV7kZXeN4
|
| |
|
|
|
|
|
| |
messages, inbox lists, calls and group chats; copies down from relay.coinos.io over NIP-77 (the strfry tunnel), never writes into it
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01QjVvE4YAcK3AGwV7kZXeN4
|
| |
|
|
|
|
| |
in deep sleep gets the wake late rather than never
Co-Authored-By: Claude Fable 5.1 <[email protected]>
|
| |
|
|
|
|
| |
justified rows
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
| |
the site card as the fallback picture (same as coinosv3 f3a3d88)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
are retired
Releasing a name disabled its own offer but left its custom offers live
(27 orphans on 2026-10-22, rick@ had 4): a payment to one found no name
and no key, and would have sat in the forward queue for good. Custom
offers now record the owner's pubkey (backfilled at start from the name,
or for an orphaned npub1xxxxxxx@ name from the one current key with that
npub prefix — 4 recovered), and the settle loop forwards by it, so a
released name's offer pays the same key's name now, as the name's own
offer already did. Orphans with no owner to follow are disabled on the
node and dropped (23), as are a released staging name's.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
was minted
A staging name is a mutinynet wallet, but every claim got a mainnet BOLT12
offer on our node published in its record, and custom offers and POS sales
were minted there too: real sats paid that way had no route onward (2,092
sat parked on 2026-10-22). otherNetwork() — the record's ark address decides,
a staging name without one is mutinynet — now gates the claim's stand-in
offer, custom offers and POS invoices (LNURL already refused staging).
At start, retireOtherNetworkOffers() disables the offers already out there
(83 staging names, 13 custom offers), takes their lno= out of the published
DNS record and drops them from state; idempotent, so a failure retries next
start. [email protected] keeps its offers: its wallet is mainnet.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
Float top-ups that missed their 60 s settle window landed later (+300k on
2026-10-22 once the ASP's vtxo pool refilled), but only a top-up seen to
settle resets the queue's backoff, so jarryd's 248k sat in a 1-hour wait
with 477k in the float.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
another network is parked
The float step threw 'address belongs to a different ark server' before the
Lightning step ever ran, so manna and rogerv3 (wallets on Second) were retried
100 times since 10-18 without once asking their wallets for an invoice. The
destination's ark id now decides: the float only on our own server, otherwise
straight to the wallet's CLINK invoice — those wait for the wallet without
paging (waitingOnWallet in /health).
A staging name's tark address paid real sats through its mainnet offer has no
route at all: such a forward is parked (state.parked, with its reason) instead
of retried forever, and /health counts it for the desk monitor. Ten such
payments (2,092 sat, Sep–Oct test payments to @staging.coinos.io names) had
been retried ~900 times each.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
top-ups, report stuck money in /health
The ASP's vtxo pool sat empty for 35 hours (2026-10-20/22: 614k in the
rounds wallet against 1.12M needed to refill), and two forwards of ~124k
for jarryd sat in the queue the whole time:
- viaClink paid the wallet's own invoice with a plain pay, but a coinos
wallet's invoice is minted by our ASP on this very node: a self-payment
the hold plugin never sees ('Destination said it doesn't know invoice').
It now goes out and back in over a channel pair (payCircular, shared with
the float top-up) and waits for the wallet to claim, so a payment is
delivered whenever the wallet is open, float or no float.
- a float top-up that doesn't settle halves the next ask (down to 20k),
back to full size after one lands: 79 tries of 200k failed where the
pool's 100k coins could have funded half of it.
- the retry tick can't overlap itself any more (a circular delivery waits
up to 4 min; an overlapping tick would forward the same payment twice).
- /health reports pendingSat, oldestPendingAt, the top-up failure streak
and last error, and payments for wallets on another ark server apart
(undeliverable): the desk float monitor pages on these.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
| |
ECC Root CA - G3, CVE-2026-111943)
names, nwcpush, blossom, the vmedia image (and live with it) and the hook's
bun install.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
| |
NIP-98), ending the stream if it's on. Its files go, and stream.json stays as 'deleted' so the address answers 410 until the trim takes it. A stream someone reported is taken aside instead (.removed/, kept for reporting), so deleting can't destroy what was reported. Removed and deleted streams no longer count as live (toward MAX_STREAMS, or as exempt from the trim).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
|
|
|
|
|
|
| |
fraction of a second behind the broadcaster where HLS is 3.5 s. The server only brokers it, and the app's secret stays here ($LIVE_CONFIG sfuAppId/sfuSecret):
- POST /live/<id>/rtc/publish (the sender, NIP-98) pushes its tracks.
- POST /live/<id>/rtc/watch hands a viewer an offer of them, and /rtc/answer takes its answer.
/stats says which publish is current (rtc), and ?watching=1 counts a WebRTC viewer, who never fetches the playlist. HLS carries on as before, for the recording and the other nostr apps.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
|
|
| |
player starts about three segments behind the edge, so most of a viewer's delay was the 2 s segments (8.6 s behind the broadcaster, measured; 4.5 s with these).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
|
|
| |
when; older records count from their last update), never if banned; GET /live/allowed says which and until when. Viewers can report a stream (POST /live/<id>/report, NIP-98): kept in .reports.jsonl and DM'd to the admins from the welcome key. An admin takes a stream down (POST /live/<id>/remove {ban}): it ends, its playlists and segments answer 410, its files move to .removed/ (kept for reporting, never trimmed), and a ban stops its sender streaming.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
| |
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
|
|
| |
(reasons.call), not to an older one whose worker would show a call as chat.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
|
|
| |
are watched; one big enough to be a call offer (an SDP) pushes 'call' at high urgency, a minute's TTL, naming the wrap, which the device fetches from GET /wrap/<id> (kept two minutes; a push can't carry it) and opens with its own key. ICE candidates and hang-ups are too small to wake anyone. A 'call' opt-out like the other reasons.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
|
|
| |
credentials good for 4 h, to a NIP-98-signed ask; 30 an hour per key, 60 per IP; STUN alone if TURN can't be had). The key comes from $STATE/live/config.json.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
|
|
| |
to the service.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
|
|
| |
WebSocket (on 443, through Cloudflare: no ports to open), ffmpeg to 2 s HLS (H.264 720p, AAC), a live window and the whole stream (its recording), viewers counted, a dropped sender resuming the same stream, NIP-98 + coinos-name owners only, 3 at once, 4 h each, recordings kept a week.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
|
|
| |
watched (mirrors coinosv3 b929771; the live container still runs ~/vmedia/server.js).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_013ZDf1cyCQLkfjftYhymZAg
|
| |
|
|
|
|
| |
the video pager's slides ahead (mirrors coinosv3 d718fce; the live container still runs ~/vmedia/server.js).
Co-Authored-By: Claude Fable 5.1 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
names/vendor/ goes.
The pin had stayed at 082fc2e (Sep 14) while the app's Ark manager moved on —
renewals priced at a 300-sat base fee and re-priced on "insufficient refresh
fee", expiry-aware coin selection, standing (scheduled) renewals merged by
unlock hash, policy-aware freeing of in-flight coins — so the registrar's
forwarder float was a month behind the wallets it forwards to. The manager's
constructor, the methods the services call (balance, sync, refresh, send,
address, createLnInvoice, driveLn, startBoard, pendingActions, htlcClaim,
htlcCosignRefund, _key) and the storage contract (load → state or null, save
← state) are unchanged between the two commits, and nothing new is imported
from a package the server lacks.
names/translate.js imports translate-split.js from the submodule now, as
feed/cache.js already did lang-guess.js: the vendored copies were byte-equal
to what the pin now carries. The README's note about them goes with them.
Resolve-checked every service entry (bun build --target=bun); the moved tests
(registry-ownership, public-feed-cache) and the bun-test files
(deepl-translate, names/store) pass.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
Claude-Session: https://claude.ai/code/session_01J1Kon8bxJoHRpJdSzshCkp
|
| |
|
|
| |
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
| |
The lockfile was written by a newer Bun than the pinned image can read. The
deploy hook now keeps its checkout as a normal clone of the deploy repo,
because submodule commands need a working tree of their own.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
| |
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
persist() rewrote the whole state file, close to 10 MB, on every settled
payment and every queue tick, so one crash during a write could take every
name with it. State now lives in SQLite (bun:sqlite, WAL,
synchronous=FULL): one row per name, offer, invoice and so on, with the
forwarder wallet's state kept as a single row. Each persist() is one
transaction that writes only the rows that changed, so a crash leaves the
previous state or the new one, never a torn file.
The in-memory state object and every persist() call stay as they were. On
the first start the JSON state file is imported, and a corrupt one stops
startup. The file is left in place, so a rollback to the JSON build still
finds its data. store.snapshot() makes a consistent copy while the service
runs, for backups.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
Carved out of coinosv3 (names, feed, nwcpush, blossom, vmedia, bridge and
their tests, with history). The shared wallet code they import (src/ark,
noffer, bolt12, popular-feed) now comes from a coinosv3 submodule pinned at
082fc2e, the commit the services have actually run on, so moving here changes
no behaviour; bumping it is its own commit. The two text helpers translation
uses arrived on the server later than that commit and are vendored under
names/vendor at the versions deployed, until the pin moves past them.
Root package.json pins the npm versions installed on the server. compose.yml
runs names, nwcpush and blossom from this checkout (read-only) with configs
and state under $STATE, outside the repo, and pins the Bun image.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
| |
The loader treated any read or parse error as a fresh registrar and started
with no names, then persisted that empty state over the real one. A crash
can leave the state file truncated, because persist() renamed the new file
into place without flushing it. Now a missing file is a fresh start, an
unreadable one stops startup, and persist() fsyncs the file and its
directory around the rename.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
| |
createConnection on a stale socket path threw ENOENT before a listener was
attached and killed the process. Create the Socket, attach .on('error'), then
connect. This was patched on the server and never committed.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
The deployed registrar carried changes made directly on the server between
Oct 10 and Oct 14 that never reached the repo:
- forward() delivers to the same key's current name when the paid name was
released in the meantime (a v2 migration swaps its throwaway npub name for
the real one seconds after a payment can settle)
- "nowhere to send" throws, so the payout stays queued instead of dropping
- the forwarder float merges small coins (consolidateFloat) before a send
hits the input cap
- floatUpkeep skips coins already past expiry, which the server's sweeper
has taken and whose refresh never settles
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
| |
server-made copy. The app shrinks a post's photo to the column's size itself (an 1100px copy made once, kept on disk) — but only when the host allows CORS; one that doesn't (primal's blossom, nostrcheck, most personal sites) left the <img> painting the original, and a 9528×6328 camera file froze a wheel scroll for 400 ms on its texture upload. vmedia now has /vmedia/pic.webp?s=col|tile&u=<picture>: the picture at 1100px wide (or a 360px square of its middle for the grid), made with libvips (shrink-on-load, so the original is never decoded whole; turned the way its EXIF says; alpha kept; stripped) as WebP, cached on disk and at the CDN for everyone. The usual guards: https on public addresses only, each redirect's host checked, 60 MB cap, the bytes sniffed (a GIF is refused — its copy would be a still), a few jobs at a time, 600 new copies per address per hour, an 8 GB cache. The client asks the server for the column copy when the host is known to keep its bytes (or has just refused, or there's no worker to make one here), keeps it in the same cache as the copies it makes, and takes the server's tile for the grid the same way; a picture the server has no copy for paints as before. The container gets libvips-tools (Dockerfile). tools/pic-proxy-test.js (canned server answers). Server measured on cs: a 3072×4608 photo in 0.5 s the first time, a millisecond after; feed-stability's two pre-existing asserts still fail on 491904e.
Co-Authored-By: Claude Fable 5.1 <[email protected]>
|
| |
|
|
|
|
|
| |
isolabellart saw the Translate button but nothing became Italian: the target was the app's language, coinos has no Italian UI (so it ran in English), and the language guess didn't know Italian. Posts now translate into the reader's choice (Settings → Nostr → Translation, any of 31 languages), else the phone's language when the app has no UI for it, else the app's; the guess knows Italian, Dutch and Polish. And, as Jumble offers: a DeepL API key of your own, kept on the device and sent per request as x-deepl-key — the registrar passes it to DeepL (api-free for ":fx" keys; DeepL refuses browsers) and never stores or logs it; a refused key or used-up quota says so once. names: translateDeepl (prose only, links/tags kept, detected source back, 401/402/422 statuses), CORS allows the header. Deployed to cs names. tools/translate-italian-test.js, tools/deepl-translate.test.js.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Vps2H27rhbG9ZqseGfMrc3
|
| |
|
|
|
|
|
| |
took a-z, 0-9 and . _ - (1–30), so bitcoin-education and a dozen others got through. A NEW claim must now match /^[a-z0-9]{2,24}$/ (v2's /^[\p{L}\p{N}]{2,24}$/u, held to ASCII since a name is a DNS label); a name already held (19 with . _ -, plus "f" and "g") is grandfathered — its owner can still update it — as are v2 migrations carrying their own name and npub-shaped defaults. /name/<n> answers invalid: true for a free name that breaks the rule; the claim form says "Letters and numbers only" as it's typed, and the suggested name from a nostr profile is stripped to letters and digits.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Vps2H27rhbG9ZqseGfMrc3
|
| |
|
|
|
|
|
| |
and a clip that fades in from black (7 of the 68 made in three hours) gave a black tile. Each frame's mean brightness now comes out of ffmpeg (signalstats YAVG on pipe:2 — /dev/stderr can't be opened from a Bun-spawned pipe), and a dark one sends it further in (2, 5, 10, 20, 40 s) for the first bright frame; a clip dark all the way through (a night scene, a boot screen) gets its brightest frame with the levels stretched (normalize). Stills are made again under a new key and URL (v=4), so neither the server's nor the CDN's black copies are served.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Vps2H27rhbG9ZqseGfMrc3
|
| |
|
|
|
|
|
| |
thumb.jpg) at once — about a second the first time anyone asks for that clip, a CDN hit after — instead of waiting on the browser's probe (64 KB) and frame grab (≈1 MB of the clip, decoded) per tile, which on a phone left most of the grid as empty boxes. The frame grab still runs for a clip the server had no picture for. A known-heavy clip wears its size badge over the still. vmedia makes 8 stills at once (was 4). tools/video-tile-poster-test.js
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Vps2H27rhbG9ZqseGfMrc3
|
| |
|
|
|
|
|
| |
translation in brackets after it — "#総合部絵チャ (…)" — the tag itself kept as the link it is. A post that is only such hashtags now counts as foreign (it was judged by its prose alone, of which it has none, so nothing ever offered to translate it), and ⋯ → Translate on a post with no words says so instead of "already in your language". The registrar's language detection reads the glossed tags too; lang-guess's scriptOf takes a minimum (a two-character tag has a script too).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Vps2H27rhbG9ZqseGfMrc3
|
| |
|
|
|
|
|
| |
from Spanish · Show original" under the words, automatic by default for posts the language guess calls foreign (Settings → Nostr → Translation switches it off and leaves a Translate link instead), and ⋯ → Translate for any post. The browser's on-device Translator does it when the language pair is already installed (desktop Chrome); otherwise the registrar's new POST /translate (names/translate.js): a cached, 4-at-a-time front for a LibreTranslate container on cs (names/compose-libretranslate.yml, v1.9.6, 32 languages; none for sw/mr/ta/te, so those readers get no offer). Links, nostr: mentions, :emoji:, #tags and long codes are never sent to the model — only the prose between them, batched in one call (src/translate-split.js, shared). Translations are remembered (coinos-translations, 300). tools/translate-test.js
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01Vps2H27rhbG9ZqseGfMrc3
|
| |
|
|
|
|
| |
when minting a terminal token (it sent none, so every terminal priced sales in USD and a $0.03 sale paid out as $0.04 CAD), and a new token retunes the terminals already paired to that name
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
| |
(the operator's own IP while testing sign-ups)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
| |
clip this browser can't picture (an iPhone's HEVC on Chrome for Linux played sound over black) — vmedia container on cs (bun + ffmpeg, /vmedia/), switch on zero-width metadata or an HEVC header from the probe, a line saying why if the copy fails too
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
/nprofile1… and /<name> with the kind 0, latest posts, follow counts and mentioned faces inlined (stale-while-refresh cache, 2 s budget), the app seeds its caches from it, and the big avatar never flashes punk art while a lookup runs
A cold visit to a profile link showed a punk, '–' counts and no posts for 9 s.
Server-side relay answers settle once events stop arriving (idle 400 ms)
instead of waiting out a relay that never sends EOSE: a cold build is ~1.9 s.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
| |
dead sockets are dropped and it retries on fresh ones (four welcomes on 2026-10-07 vanished after the uplink dropped while the log said sent)
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
68fd614 removed the VtxoPolicy::Htlc encoding from send.js/proto.js (upstream
took type 0x08) and the bridge was retired from the app on 2026-08-02, so
htlcLock can no longer build its output (policyTaproot throws on the missing
userPubkey). bridge-e2e-test also drove bridge/server.js, which sits on the
same dead primitives.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
| |
coinos.io's v3 bridge forwards a migrated legacy account's incoming
payments to the v3 wallet only on proof: the holder is the legacy
account's own key, or the wallet its session named via /migrate-grant
(granted: true).
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_011bvzKaZosXGHkigcGcSzAR
|
| |
|
|
|
|
|
|
|
| |
Android holds 'normal' pushes while the phone dozes and drops them once
the 1 h TTL runs out; a pocketed phone missed most zaps and mentions
that the laptop showed. Reactions and community chat stay normal.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_011bvzKaZosXGHkigcGcSzAR
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The server renders /nevent1 and /note1 pages with the thread inlined
(post, root, replies, kind 0s) and Open Graph tags; the client seeds its
thread cache from it and opens the thread on the first paint. Without
it, a pending frame holds the screen instead of the front door. Profile
lookups ask our relays and the profile indexes at once, streamed,
instead of one after the other behind five-second waits. A thread
opened straight from a link has no Back button.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
|
| |
|
|
|
|
|
|
|
|
|
|
| |
The notifier now watches kind-7 reactions p-tagged at the user (opt-in
by an explicit flag, so older apps that can't word them stay quiet) and
sends zap receipts as 'zap' to apps that report the category, 'payment'
to older ones. Coinos zap receipts carry their amount into the push. The
service worker names the reactor when it knows them and stays quiet for
a muted one.
Co-Authored-By: Claude Opus 5.5 <[email protected]>
Claude-Session: https://claude.ai/code/session_01VP5Ljqk7EYcm5QQAbzDwog
|
| |
|
|
| |
Build the default Popular feed once on the server and let browsers verify and paint its signed events before slow media finishes. Keep custom relays, pagination, filtering, and relay fallback on their existing paths.\n\nAdd stale-while-refresh persistence, service-worker bypass, development integration, and unit/browser coverage.
|
