Search: GET /api/search sent each matching user's full row — password hash and email included — to anyone, signed in or not, because searchUsersGlobal returned User[] and the route passed it through. It returns profiles only now, stripped where the rows are read, and the search test fails on the old code. Types alone wouldn't have caught it (a User is structurally a PublicUser), so the strip is explicit and tested.
Typed routes. Every JSON response is declared as APIResponse<T>, the way the typed routes already did, with the missing shapes named in api/model/interfaces.ts: PublicUser (no hash, no email) and SelfUser (your own profile, with email) plus toPublicUser/toSelfUser in place of ten hand-written destructurings, AuthTokens, AuthResponse, MediaTokenResponse, InviteResponse, AttachmentResponse, SearchResults, LinkPreview, UserSuggestion and Invitee. Request bodies are `unknown` and narrowed by the existing guards, or by a new fieldOf(obj, key) where a route reads one field — no more reading off Oak's `any` or casting to Record<string, unknown>. The password-reset routes answer `{ success: true, data: null }` rather than a bare `{ success: true }`, so every success fits the envelope; the frontend only reads `success`.
Socket protocol: routes/ws.ts had its own send() taking Record<string, unknown>, so nothing checked what it sent. It takes ServerToClientMessage now, which turned up two mismatches: `pong` was missing from the union although the server answers every client ping with one, and vote_ack carried a `success: true` no type declares and no client reads. The union has PongMessage and vote_ack lost the field.
No explicit any. The YouTube playlist parse reads ytInitialData as `unknown` through a small at(node, …path) that checks each step, instead of `type Json = any`. The backend test helpers type the response body as the server's envelope: request<T>() and get/post/put/patch/del<T>() name what `data` holds, res.data throws with the server's message when the call didn't succeed and res.error when it didn't fail, and each test says which API type it expects, so a test reading a field the route doesn't send no longer compiles. Assertions that a field is absent are `assertFalse("passwordHash" in …)` rather than reading a field the type says isn't there. The socket tests' messages are ServerToClientMessage, with next()/all() keyed by type so each returns that message's shape.
No double casts in src/ either. window.YT and onYouTubeIframeAPIReady are declared on Window; the test's fake player is a class forwarding to the spies the tests inspect. WSContext, FollowContext and PlayerContext export their defaults (defaultWSValue, defaultFollowValue, defaultPlayerValue), which the tests import instead of reading React's private _currentValue. handleMentionKeyDown takes a MentionKey — the key fields it reads — so the test builds a plain object; useUserDumpFeed declares sentinelRef as the callback ref it is. In src/test/setup.ts the IntersectionObserver fake implements IntersectionObserver, and report(isIntersecting, top?) builds complete entries in place of partial ones cast in five files; the Animation fake implements Animation; the canvas getContext stub goes in through defineProperty. The XHR fake's last request is optional behind a getter that fails if none was made, the keyframes are checked with matchers, and the roleless legacy session is built by leaving the role out. One test changes meaning slightly: "leaving the view…" relied on entries with no position at all, and now leaves and re-enters at the same height, which is what it describes. The one cast left is in api/test/ws.test.ts: deno.json's "dom" lib hides Deno's WebSocket overload that takes handshake headers, and dropping that lib would break the editor's view of src/; it's commented and names the real signature.
Also: the notifications test answered a Response-typed promise with a plain object, which failed `tsc -b`; it uses the fetch stub's deferred() now. ESLint didn't read the `deno-lint-ignore no-explicit-any` comments, so `eslint .` failed on three files; those anys are gone. `deno fmt` over api/ and src/ (not the compiled src/locales/*.js); the reflow left all 332 message ids intact, and the .po files only have their source line references updated.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Fixed along the way. Dumps: clearing a dump's note sent `comment: undefined`, which JSON drops, so the server kept the old one; it's sent as null now. Replacing a file dropped a visibility change made in the same save (the file endpoint doesn't read it); it follows as a separate update. Renaming a dump onto a new slug kept showing the old title until reload. The thumbnail picker and delete on the edit page failed without a word, the latter as an unhandled rejection. Playlists: an edit that kept the slug waited on the socket's echo to show, and a new cover never appeared (the upload's answer was ignored, a refused one said nothing); covers were also served stale for an hour, since `max-age=3600` met an unversioned URL — setPlaylistImage now bumps updated_at and every cover URL carries `?v=` (playlistImageUrl). Removing a dump, reordering, and deleting a playlist ignored non-OK answers, leaving page and server apart; they restore, reload or report. Profile: clearing "Who am I?" sent undefined and came back on reload; your own upvoted list (and the upvoted page) was empty until the socket's welcome, and forever without one; "View all" under Following linked to a route that doesn't exist and is a Load more now. Chat: Escape on "Delete this message?" closed the whole chat — ConfirmModal takes Escape first, in the capture phase. Player: a pause or play from a media key, headset or the OS wasn't reflected (MediaPlayer follows the element's own play/pause events), and the seek bar tracked dragging with mouse events only, so on touch the thumb fought timeupdate. Feeds: a guest on /~/followed got a blank page; a server error in the followed feed read as "Could not connect". Search: a failed later page threw away the results shown. Notifications: a failed mark-read replaced a loaded list with an error. Silent failures now say so: deleting a comment, the voters/likers "Load more" (stuck on "Loading…"), and the password-reset request, which said "sent" whatever happened.
Notifications also stop hiding what was new: the bell still clears on opening the page, but the dots and the "N new" count stay for the visit instead of vanishing the moment the mark-read lands, and one arriving while the page is open is marked read too, so the bell doesn't count it.
Accessibility: Modal and ConfirmModal are dialogs (role, aria-modal, labelled/described); playlist membership rows are checkboxes that work from the keyboard; a dump's title, and the profile's email and description, can be opened for editing from the keyboard (the pencil shows on focus); the category editor's ✓ and the global player's video button have names. Player, rich-content and download labels were hard-coded English and are translated, with the new French strings. Lint: fast-refresh and render-purity rules are off for src/test/ only, where probes hand their context out on purpose.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Two things set the rhythm, and neither is a row counter — that beats like a metronome. It comes due on a count of *cards*: a block is a set of stories and a hero is one story however many cells it takes, so a stretch thick with big footprints reaches the count over more rows than a stretch of squares does, and block heights end up tracking the size mix (measured: four to five rows through square-heavy runs, six to seven where art clusters). And the break *makes* its own section front rather than waiting for one, promoting the card that opens a new block to a hero. Waiting was the first attempt and it failed on the numbers — a banner lands at the start of a row on its own two to five times in thirty-six, so the preference almost never fired and a max-rows cap did all the work. Promotion always fits, since a break only falls on a row with all four cells free. A tail guard keeps the last cut from leaving a rule with a fragment under it.
The break sits between two grids rather than inside one: a grid item would be stretched to the 9rem row track, and out here its height is its own, down to nothing. Three custom properties are the whole contract a theme writes against — the air around it, the rule through it, an optional glyph centred on that rule — with `none` on the last two meaning the real absence of each, so a theme can remove the separation entirely rather than leave an empty band. Themes wanting something a rule can't be reach past the properties to plain CSS: nyt pairs hairlines in the grid's own pen and clips them to the art inset (which moves up to the feed, since a property on the grid can't reach a sibling); brutalist spends air around the 2px border it already divides everything else with; kanagawa takes a comment divider, geocities the rainbow its header ends on, melange a thinner band of the mélange. The page gutter moves to a feed wrapper so every block lines up on the same edges, and nyt's masking frame now runs once per block, which is what makes each of them read frameless and not just the feed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The emoji picker and the @-mention list were absolutely positioned inside
.mention-textarea-wrap, so every ancestor with a clipping box cut them off —
and TextEditor sits inside plenty: .dump-card-inner, .journal-card and
.rich-content-card clip with overflow: hidden, while .modal-body and the chat
scroller clip just as hard with overflow-y: auto. The picker also always
opened upward, so a textarea near the top of the viewport pushed it off-screen.
Both popups are now portaled to <body> and anchored to the textarea with
position: fixed, which no ancestor can clip. The new useAnchoredFloat hook
flips to whichever side has more room, clamps horizontally on narrow screens,
and publishes the space left as --float-max-h so the list shrinks and scrolls
instead of overflowing. It re-places on scroll, on resize, and through a
ResizeObserver, since autoResize grows the textarea as you type.
z-index goes to 1100 to clear the modal backdrop, staying under tooltips.
Drops two point-fixes that only existed to work around the clipping: the
hardcoded upward flip for the chat composer's mention list, whose leftover
margin would now double the gap, and the claim that the chat modal's body
needed overflow: visible for the popovers' sake.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The dump card is proportioned for the 860px feed; in a ~400px profile
column the title wrapped four lines deep, the meta row stacked, every
category chip took a line of its own and the vote button floated in the
middle of the resulting void. Re-proportion it for the column — smaller
thumb, title clamped to two lines, no teaser, meta and chips sharing the
line under the title — scoped to the width where the columns split.
Also stop the playlist card's inner from clipping its own hover: it
inherits overflow:hidden from the rule it shares with the dump card,
which needs it for the collapse animation. A playlist card never
collapses and its height is set by its 80px square thumb, so the clip
shaved the corners off the scaled thumb and sliced its lift shadow into
a hard-edged slab against the card border.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Posting a dump was a single form where the important choices were the easiest
to miss. It is now three panels: link or file, why & where, playlists.
Composition:
- No more URL/File toggle. An empty panel offers both ways in at once and the
kind follows what you actually did; a file dropped anywhere in the modal is
accepted, not just on the zone.
- Categories and visibility get their own panel instead of a disclosure that
read as optional, and the primary button stays "Next" until they've been
seen. Visibility carries a real label now.
- The draft (link, title, why, categories, visibility) is mirrored to
localStorage on every change and restored on reopen, so Escape or a stray
backdrop click costs nothing. Only an attached file can't be restored, so
that is the one case that asks before closing.
- URL dumps can carry a poster-supplied title instead of being stuck with
whatever the page scraped, editable right under the preview.
- Multipart uploads go through XHR so there is a real progress bar and a
percentage on the button, rather than 50 MB of silence.
Duplicates:
- New dumps.url_canonical column (+ index, backfilled by 0013) holding a lossy
key that ignores scheme, www., trailing slashes, tracking parameters and
YouTube share shapes. GET /api/dumps/by-url reads it, and the create form
warns "already dumped by X" while it fetches the preview. Never blocking.
Fixes:
- /api/preview now reports whether the page was actually reached: a failed
fetch still yields a hostname-only stub, so a dead link and a page without
metadata used to render identically.
- The Web Share Target never worked. The manifest posts to "/", but the index
redirect dropped the query string, so every Android share landed on the feed
with nothing pre-filled.
- File dumps no longer take the extension into their title.
- The link field no longer autofocuses on touch, where it raised a keyboard
over the modal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TiAPtJZeCYYk8rKehtLUQU
The dump modal's window-level paste listener looked at the clipboard files
before it looked at the event target, so an image pasted into the "Why?"
field was claimed twice: the editor uploaded it as an inline attachment,
and the modal switched itself to file mode with that same image as the
dump. Writing a URL dump with an illustrated description was impossible.
The target guard that already existed for text pastes now runs first and
covers files too — fields own their paste, and the URL input keeps its own
handler for the image-to-file-dump shortcut. Pasting onto the modal body,
where nothing else is listening, still starts a file dump.
Card teasers are line-clamped boxes, which an embedded image blows apart.
Inline markdown — the mode the dump and journal cards already ask for —
now drops img nodes, so a description image shows up on the detail page
and nowhere else.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01TiAPtJZeCYYk8rKehtLUQU
Re-resolving a stream is a network round-trip, and nothing checked that the
player still belonged to that item once it landed. Picking another dump —
or closing the player — while a resolve was in flight let the stale result
swap the queue back; from onError, which resolves with autoplay, it would
also start playing over whatever was chosen instead. A generation counter,
bumped by playQueue/advanceTo/stop, now drops any result that no longer owns
the player, with a separate counter owning the resolving flag so a
superseded resolve can't clear a newer one's spinner.
A track that stays dead after a re-resolve is one track, not one album: the
cooldown path now steps to the next entry and only falls back to the iframe
on the last one. A failed resolve still goes straight to the embed, since
that's an album-wide failure rather than a single bad URL. And a track that
has vanished from the release, or lost its streamable flag to Bandcamp's
free-play cap, no longer resumes track 1 at the dead track's offset — the
findIndex miss resets the offset instead of seeking past the end of another
track.
Clicking the playing row rewound it but never resumed: seekTo only moves
currentTime, so re-selecting a finished or paused track looked like a
no-op. It now resumes as well. Same idea in the compact rich-content card,
whose button advertises "Pause" while active but re-played on click — in
native mode that meant a fresh /api/bandcamp/tracks and a restart from
track 1. It pauses for streams now; embeds have no transport of their own
and keep restarting, as before.
playRichContent reports whether playback actually started, so a native-mode
Bandcamp page with no streams and no stored embedUrl — an artist root, a
/music index, a preorder — is no longer a dead click: the journal card
navigates to the dump as it used to, and the rich-content card opens the
source page.
Also: the tralbum fetch releases the response body on its two error paths.
The endpoint is unauthenticated, so random /track/<slug> URLs would leak a
connection per request while also defeating the cache.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01E56F55N7m5GGKKEwoYFJtH
The player now holds a queue instead of a single item, so albums play
through and single tracks share the same code path. Adds a "stream" item
kind for expiring, remotely-hosted sources: Bandcamp signs its mp3 URLs
for ~24h, so an item carries the page it came from and re-resolves itself
on error or on a stale restored session, falling back to the iframe embed
if that fails too.
Gated behind GERBEUR_BANDCAMP_PLAYER (default "embed"), injected into a
meta tag the same way as site-name/site-emoji.
Also: player header gains artwork and a subtitle, volume persists across
queue advances and reloads, and cross-origin streams get a plain seek bar
rather than a waveform that can never decode.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SbtjNsT5wvuABnfqhegZEJ
