Repository navigation
Conversation
Overnight, an entire publisher's comics would lose their read state. The files were fine; the rows were destroyed and re-imported with new pks, and Bookmark.comic cascades. Three layers in sequence each read "I could not look" as "it is not there", and a fourth turned one failed stat into a whole publisher. The disk walk drops an unreadable subtree with no record that it did — silently, in the per-entry branch. The diff is a bare set subtraction, so every database path under that subtree reads as deleted. The delete backstop confirms the lie, because Path.exists() swallows every OSError and answers False for a file it merely could not reach. And the folder delete cascades its comics away without a stat of its own. Only ENOENT and ENOTDIR now mean gone. Every other error means the filesystem could not answer, which is not the same thing: - Snapshot carries an `unreadable` set. Both walk failure sites record into it, and the silent one finally logs. - SnapshotDiff withholds deletes for those paths and everything under them, and says how many. A withheld row still points at a real file, so the next healthy poll reconciles it. A file carried out of the subtree can still pair as a move, which is what keeps its bookmarks. - split_extant stats instead of Path.exists() and fails closed, naming the errors it could not check so a recurrence is diagnosable. - bulk_folders_deleted probes the comics its cascade would take, drops every folder above one that is still on disk, and calls the mass-delete warning that was written but never wired to it. Every cascade now logs how many comics it is about to remove. What remains unhandled is an outage that answers ENOENT for files that exist. The new logging is how that would be recognized. Tests mock scandir and stat rather than using chmod: CI runs as root, where mode bits are ignored, and a 0o000 directory still stats from its parent, so a permissions test can only reach one of the two routes. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* fix(opds): restore library browsing on the v1.2 start page Panels on macOS showed Keep Reading, Latest Unread and Oldest Unread and no way into the library (#855). The four browse views — Publishers, Series, Folder, Story Arc — were emitted as entries only when the client was judged facet-blind, and as feed-level facet links otherwise. macOS Panels ignores facet links, so its start feed was a dead end. Two faults, fixed separately. Navigation was folded into the facet mechanism. A facet refines the entries of the feed it appears on, and the start page has none by construction, so "Views" refines nothing: it is navigation wearing a facet costume. Failing to render facets may cost a client its sort UI, which is cosmetic. It must never cost the only way into the library. ``facets()`` now returns sort facets only and ``nav_views()`` emits the Views group as entries for every client, called where the old facet-entry call sat so the bytes are unchanged for clients that were already getting entries. Facet-aware clients lose "Views" from their native filter menu and gain four navigation rows, which is strictly more discoverable. The Panels facet gate could not work in either direction. #817 replaced a build floor with a denylist of facet-blind builds, guessing 951 for macOS; the real builds are macOS 957 and iOS 956, one apart and interleaved, with identical Darwin tokens. No floor separates them, the guess was wrong, and a denylist or allowlist grows by one entry per Panels release forever. So codex stops choosing: Panels gets the facet links *and* the navigation folder sort. iOS users see five sort rows they do not need; macOS users get a sort instead of none. Nothing reads a build number any more, so the parse goes too. Also records, in the code, why the Folder and Story Arc entries advertise kind=acquisition and why that was left alone after checking Panels 957 — it has now been re-derived twice from scratch. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * fix(opds): make the OPDS urls dialog show something, always Second half of #855: "when I click on the OPDS link in the browser the window goes inactive but no new window appears." The dialog body was a bare div, so the overlay was a shrink-to-fit box, and the only thing in it while the urls load is a placeholder that hard-codes height and width at 50%. Percentages against a shrink-to-fit parent resolve to nothing, so the dialog opened as an invisible sliver behind the scrim — inactive window, no visible window. A v-card gives the body an intrinsic size and the placeholder is sized explicitly at this one call site, which is the only bare usage in the tree. A failed request was also unrecoverable: loadOPDSURLs ended in .catch(console.error), leaving opdsURLs undefined forever with nothing to render and no way to retry. It now records why and the dialog offers a retry. Clipboard rows were dead on every plain-http LAN deployment for a related reason: navigator.clipboard is undefined on insecure origins, so useClipboard reported unsupported, which hid the icon, suppressed the tooltip and made the click a no-op. The legacy fallback restores all three. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * news: opds browsing, panels sort, and the opds dialog Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Every guard in #858 keys on the filesystem raising an error. A share that answers "no such file" for a file that exists defeats all of them, because not-found is exactly what a real deletion looks like. A CIFS reconnect window does that for a few seconds at a time, and so does an external rm + mv caught mid-swap. One observation cannot tell those apart from a deletion. Two can, for the short shape: the delete phase now probes the paths a scan reported missing, waits five seconds, and re-probes only the ones that came back gone. Anything that answers the second time was never gone, and is dropped from the batch with a warning naming the paths — a flapping share is worth saying out loud. A scan that found nothing missing never waits, which is the common case. The wait is spent on abort_event.wait, so an admin abort cuts it short; daemon shutdown does not set that event, so a shutdown landing inside the wait costs at most one delay. Covers are excluded because they never come from a scan. This does not cover a sustained outage, where both probes answer gone. That needs confirmation across polls and stays unbuilt. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
…me (#862) An inode is identity, so it pairs a move outright. Plenty of moves do not keep one: a copy-then-delete, a cross-device mv, a remount that rotates the inode space. The poller's own docstring admitted the cost — "legitimate cross-remount renames still degrade to delete+add (and the user loses the comic.pk's bookmarks)" — because the delete cascades Bookmark.comic away and the add re-imports the same file as a fresh, unread comic. The existence backstop cannot help: the old path really is gone. Move detection now has a second tier. What the inode loops leave unpaired is matched on a signature of filename, size and mtime, and only when that signature is unique on both sides, so a wrong pair is impossible by construction rather than unlikely. The filename carries most of that weight: a bulk copy hands many files the same size and mtime, and without the name they would pair with each other. The cost of keeping it is that this pairs a file that moved, not one renamed to a different name; that case has nothing left to match on. Files only. A directory's size is meaningless and its mtime moves with any child change, so children pair one at a time and the importer creates the destination folders. Withheld paths are not indexed either: an inode may still pair one because an inode is identity, but a signature is a guess and an unreadable subtree may simply come back. mtime compares exactly, as _is_stats_equal does, so a filesystem that truncates it across a copy falls back to today's behavior. Fixed alongside, because every paired row would have hit it: the stale-stat refresh stored the mtime as an int. list(os.stat_result) takes the sequence form, whose mtime is truncated, while set_stat stores the float — and that model method warns about this exact trap in a comment. A refreshed row therefore disagreed with its own disk snapshot on the next poll and was re-imported once for nothing. The watcher pairs by inode only as well. Porting this there is a separate change: an add inside a 60s watcher batch may be an in-flight copy whose size and mtime are not final. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
The scanner matches comic suffixes with re.IGNORECASE and the importer stores the path verbatim, so Foo.CBZ is a perfectly ordinary comic. The nightly integrity check compared those same suffixes case-sensitively. An uppercase comic therefore contributed no folder to the set the check considers needed, and was skipped by the repoint step as well. The prune then deleted its folder as empty, with no existence probe of its own, and Folder.parent_folder, Comic.parent_folder and Bookmark.comic all cascade: the folder took the comic, and the comic took the read progress. A library whose comics all carry uppercase suffixes loses its root folder row and everything under it, every night. The suffix set now derives from comicbox's FileTypeEnum instead of a hand-written tuple, so a new archive format cannot be left behind here again, and the comparison folds case. Not filters.match_comic, which consults unrar and pdf support: on a host without unrar every .cbr row would drop out and be pruned, the same bug from the other side. The prune also gets a backstop that does not depend on that rule being right: a folder any comic still calls its parent is never stale. The comment claiming the repoint step guaranteed as much was false for exactly the rows this fixes; now it is true. Installs that already ran migration 0043 with uppercase comics lost those bookmarks then. A user-data sidecar from before the loss may still hold them; restoring one is manual and not part of this. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
**The blind choice.** Two candidates can read identically — same series, issue and year — and the dialog showed nothing else. The reporter's screenshot has 92% and 84% with no way to tell which is which. Each row now carries the cover, a link to the source's own page, the volume ordinal where the source has one, and what the blended score is made of: "Match 100% · Cover 60%" against "Match 100% · Cover 20%" says the cover comparison *is* the whole difference. Covers load straight from the two CDNs. An <img> is exempt from CORS, so nothing on their side had to cooperate; codex's own img-src was the only thing in the way, and this file already relaxes it for unpkg's pdfjs images a few lines up. Two path-scoped prefixes is the whole backend change — no proxy view, no cache, no SSRF ladder to contain a server-side fetch of a url that arrived in a third-party API response. **The clipped Pause button.** Vuetify's .v-card-title is nowrap with overflow:hidden, so as a flex container it clips its last child: the buttons cannot shrink below their content and the title will not yield. Two utility classes fix it — the title truncates, the buttons do not shrink — with no override and no layer question. **"already exists" on a converted CBR.** Writing tags to a CBR repacks it as a CBZ. With delete-original off the CBR stays beside its own output, and every later write collides with it. comicbox refused, but only after opening, merging and serializing, and its message named a filename with no hint what to do; because tag-write errors dedupe by path, that vaguer text also replaced the clearer one the planner had recorded. Codex now refuses first, so there is one message per path and it says to edit the twin instead. With rename on it checks the scheme name too: the kept conversion is renamed after the write, which freed the plain destination and let every retry mint another copy. Which archives convert now comes from each row's file_type — comicbox's own sniff, recorded at import — rather than the filename, so a CBZ misnamed .cbr is written in place instead of mis-planned. One constant serves the preflight view and the writer. **The missing Failed Imports link.** It was gated on *unseen* failures, so "Clear Warning" hid the only way back to a table that was still there. It now persists while failures exist and is colored only while unseen; the hamburger dot stays on unseen, because a dot is a notification. The Tag Write Errors hint says a damaged archive is listed there, and that Failed Imports is a different set — a file that imported fine and later failed a write will not be in it. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Its docstring says "Delete bookmarks without users or sessions", and it has never deleted one. The filter also required comic to be NULL, and Bookmark.comic is a non-nullable foreign key, so the query matched nothing for its whole life. The comic term was never needed either: that FK cascades, so a deleted comic takes its bookmarks with it and cannot leave one behind. What can be left behind is a row owned by neither a user nor a session — invisible to every reader and returned by no query. No ORM path creates one: the auth filter is exclusive and non-null, restore forces a user, and cascade_if_user_null deletes session-owned rows whose user is NULL while nulling the session on the rest. So a both-NULL row comes from raw SQL, a migration or legacy data, which is what a backstop sweep is for. The correct filter is the one cleanup_orphan_settings two lines below already uses; this was a copy of it with one relation too many. Because the sweep is about to delete rows for the first time on every install that has ever run it, it now names the pks it is about to remove before removing them. Checked against two real databases first: the reporter's copy has 756 bookmarks and zero orphans, and this dev database has none at all. Turning the sweep live is a no-op on both. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Writing tags to a CBR repacks it as a CBZ at a new path. With "delete original" off the CBR stays beside its own conversion, and every later write is refused. The message says to edit that comic instead, but left the admin to go find it. The refusal already queries the twin row to know it exists; take its pk and library path from the same query and carry them into the tag-write error record, so the Tagging panel can link to the comic and the log line can name it. The two keys are optional and written only when there is a twin, so every other failure keeps the shape it has today and records already sitting in the filesystem cache still read. Also adds the NEWS lines that #861, #862, #864 and the fixes half of #863 shipped without. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
``Bookmark.unique_together = (user, session, comic)`` never enforced anything. Both owner columns are nullable, and SQL treats NULLs in a unique index as distinct, so (user=1, session=NULL, comic=5) could be stored any number of times. The writer selects before it inserts, but nothing serializes those two steps: a queued page turn and a synchronous mark-read for the same comic both find nothing and both insert. The duplicate is not cosmetic. Comic cards sum page and finished over the rows, so a collection can report more finished children than it has; the reader takes whichever row it read last; and OPDS progression and the settings restore both fetch a single row and raise MultipleObjectsReturned, which surfaces as a 500. Two partial unique constraints say what was meant, each conditioned on its own owner being present. Migration 0054 merges the rows that already violate them first: the most recently updated row of a group survives and takes the group's progress with it. sqlmigrate shows DROP INDEX plus two CREATE UNIQUE INDEX ... WHERE, so no table is rebuilt. The write path now runs both passes in one transaction and catches the IntegrityError the loser of the race gets, re-running the update pass for the row the winner just inserted. Without it the bookmark thread would swallow the error and lose the page write. Verified against a copy of the reporter's database: 756 bookmarks, one duplicate group, 755 after, the surviving row still finished. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
The register form read `this.MIN_PASSWORD_LEN`, and the auth store has only `MIN_PASSWORD_LENGTH`. The read was `undefined`, `v.length < undefined` is always false, and the whole branch had never run: any password of one character registered. Had it run, its message would have read "Password must be undefined characters long". Fix the identifier in both the mapState key and the rule, and use the wording the profile and reset-password dialogs already use. The new spec seeds the store with a minimum of 6, not the default 4, so neither a misspelled store key nor a hardcoded literal can pass it. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Vue 3 does not invoke a mixin's setup(), so the global keyup listener in auth-form-mixin.js has never executed in any of its four consumers. Making it run would be worse than deleting it. These dialogs mount whenever their parent renders, not only when they are open — auth-menu renders ProfileDialog or AuthLoginDialog unconditionally, and login-dialog supplies its own activator — so a component-lifetime listener calling stopImmediatePropagation would kill keyup app-wide for the whole session. The protection it was meant to provide already exists and works: all three reader keyup handlers bail on the auth store's isAuthDialogOpen getter. Verified in the browser — `m` opens the metadata dialog from the reader, and typing "mmm" into the login dialog opened from the reader does not. @vueuse/core stays in package.json; the reader components, the socket store and clipboard.vue still use it. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
The edit panel binds :rules to five fields, but grep -c v-form found zero: nothing ever called validate(), so canSave ignored every rule, and TagWriteRequestSerializer.patch is an opaque CharField so the backend never saw them either. buildPatch clamped out-of-range values on the way out — the author left a comment saying exactly that — so a year of 10000 was silently written as 9999 and the user was never told. Follow the admin tabs' precedent: a <v-form ref="form"> root and a validate() early return at the top of preSave. The Save button stays enabled; clicking it validates, renders the messages and aborts. @submit.prevent guards against VForm's own onSubmit calling the native form.submit(), which would be a full page navigation. VForm.validate() iterates every registered input and never checks disabled, so a stale value in a field the selected formats cannot write — a rating count with no average under COMIC_INFO — would block an unrelated save with an error on a field the user cannot reach. activeRules() gives a disabled field no rules. The root element is replaced rather than wrapped: wrapping re-indents 1370 template lines. #editPanel's styles have no child selectors and nothing else references it; #editToolbar's sticky containing block is still the same content box. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
…870) * refactor(rules): hoist static validation rules to module constants Nine rule arrays were computeds that never read `this`. Frozen module constants exposed through data() discharge the "no computed for static values" convention and shrink the plugin conversion that follows to one line per rule. Vue leaves a frozen array raw, so binding it from data() costs nothing. No template changes: each data() key keeps the name the template already binds. auth-tab keeps URL_REGEX — canEnableOidc uses it independently — and profile-dialog keeps passwordRules a computed, since it reads passwordSectionActive. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(rules): install Vuetify's rules plugin with Codex's aliases createVuetify() does not install the rules plugin — it is a separate Vue plugin that needs the locale instance createVuetify() returns. One composite default export keeps `app.use(vuetify)` the single install point for main.js and for all 41 specs that pass this export in global.plugins, with no second install path to forget. That matters because an unresolved ["$alias", …] array fails *open*: Vuetify warns and skips it, and the field validates with no rules at all. rules-plugin.test.js pins that behavior from both directions — no plugin, and a misspelled alias — and scans every $alias used under src/components against the registry. Three aliases: $intRange and $numRange share one range builder, differing only in Number.isInteger vs Number.isFinite, and treat blank as unset because every numeric field in Codex is optional. $notIn unwraps a getter at validation time, so a set that reloads on a WebSocket event while the dialog is open is read fresh — Vuetify does not re-validate merely because the rules array changed. throttling-section is the canary; its message is already the alias default, so it needs no message argument. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * refactor(rules): express integer ranges through the $intRange alias Five hand-written range closures become alias descriptors: the three publish-date parts, the SMTP port and timeout, and the custom-cover max upload. DATE_PART_BOUNDS and DATE_PARTS stay — buildPatch clamps with the bounds independently, and edit-panel-date.test.js asserts that clamp. custom-covers is the one site with no blank guard: Number(null) and Number("") are 0, which is below the minimum, so blank fails today with "Must be 1–2048". $intRange alone would pass blank, so $required runs first carrying the same message. Every input gives the same result as before, including 0, which $required lets through by its own carve-out and $intRange then rejects. "the rules bound each part" called each rule as a function; a descriptor is not callable. It now drives the real fields through validate(), which is what the descriptors mean anyway. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * refactor(rules): express rating range and name uniqueness through aliases The community rating becomes $numRange — same builder as $intRange with Number.isFinite instead of Number.isInteger — and the two "already used" checks become $notIn with a getter. The getter is required, not stylistic: usernames and names read admin store tables that reload on WebSocket events while the dialog is open, and Vuetify does not re-validate because a rules array changed. Passing the Set directly would freeze whatever it held when the dialog opened. Two deliberate deltas, both invisible: blank now passes $notIn, with the required rule ahead of it rendering the same message, and nameSet() returning false for a non-admin no longer throws a TypeError inside validation. communityRatingCountRules stays a closure — it reads the sibling community_rating out of patch. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* fix(style): correct two rbg( typos that silently voided a colour
An invalid function name makes the whole declaration invalid, so both
sites have been falling back to the inherited colour. Neither is a
visible change — the fallback happened to be close enough that nobody
noticed — but the declarations now do what they say.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(browser): show the card hover border that never rendered
`border: solid thin` on the line above has always had its colour
discarded, because `rbg(` is not a function and an invalid value voids
the declaration. The border fell back to `currentColor`.
Unlike the other two typos this one is visible: hovering a card now
draws the orange border its author intended.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* chore(style): delete dead theme config, dead CSS and a dead dependency
- `options: { customProperties: true }` is a Vuetify 2 key; Vuetify 3
emits the --v-theme-* variables unconditionally.
- `new createVuetify(...)` — createVuetify is a plain factory. `new` on
a function returning an object discards the constructed `this`.
- The CSP comment named `theme.isDisabled`, which does not exist (the
real switch is `theme: false`), and its goal is unreachable anyway:
Django already ships style-src 'self' 'unsafe-inline' for Vue's
scoped styles. Rewritten to say that.
- anchors.scss's commented `a:visited` block references
--v-theme-secondary, a colour the palette does not define.
- pager-vertical.vue's debug CSS comment.
- browser-table.vue hardcoded the 0.12 that Vuetify emits as
--v-border-opacity; same value, now sourced from the theme.
- typeface-roboto was in package.json with no import anywhere.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* test(theme): pin the theme and style contract
Nothing about the theme was tested. A rename, a stray hex in `defaults`
or a typo'd colour function changed rendering with no signal — which is
how `rbg(` survived in three files, one of them hiding a card hover
border for years.
Two contracts, no mounting and no new dependencies: what
src/plugins/vuetify.js declares, and what the style blocks may say. The
scan asserts it found files, so an empty list can never make the rest
vacuous.
Verified it bites: altering one colour hex and typing `rbg(` into a
component each fail their test by name.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* refactor(theme): pass theme tokens to defaults, not resolved hex
Ten component defaults handed Vuetify `codexTheme.colors.primary`, the
literal "#CC7B19". A hex lands as an inline style on the element; a
token resolves to a theme class at render time, which is what every
other colour in the app does and what keeps a colour change in one
place. reader-toolbar-nav's :track-color and the edit panel's disabled
Save button did the same thing by other routes.
The contract test gains the assertion, so a hex cannot come back in.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* refactor(theme): kebab-case the eight custom colour tokens
Vuetify's own tokens are kebab-case (surface-light, primary-darken-1,
on-surface); Codex's eight were camelCase, so every style block mixed
two spellings of the same idea. Values are unchanged — the four text
greys keep their exact hexes and DESIGN.md §3 survives verbatim, only
the names in the table move.
The auto-generated utility classes become .text-text-secondary and the
like. Nothing uses them, and the edit panel's disabled Save button emits
text-text-disabled, which is fine.
The contract test gains a kebab-case assertion: a reintroduced camelCase
variable resolves to nothing and the declaration is dropped in silence.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(admin): give the tagging validation chip a real icon, drop @mdi/font
The chip passed the class names "mdi-check-circle" / "mdi-alert-circle"
to :prepend-icon. Codex runs the mdi-svg icon set, which resolves an
icon name to an SVG path; a font class name is not one, so the chip has
rendered an empty icon slot. It now passes the paths from @mdi/js, as
every other icon in the app does.
That was the last consumer of the webfont, which main.js imported
globally. All three greps are empty — no `class="... mdi ..."`, no
"mdi-name" strings under src, nothing in the Django templates — so the
import and the dependency go.
main-*.css drops from 508 KB to 192 KB, and no materialdesignicons
webfont files ship at all.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* fix(pwa): correct the status-bar meta and the manifest theme colour
apple-mobile-web-app-status-bar-style accepts only default, black or
black-translucent. It held "#2A2A2A", which iOS ignores, so the setting
has never applied. black-translucent gives the dark chrome the hex was
reaching for and lets the app paint under it.
The manifest's theme_color was the codex orange. theme_color is the app
bar, not an accent, so an installed Codex had an orange bar over a dark
app. It is surface-light now, matching the theme-color meta and the
background_color beside it.
Both sites carry a comment naming src/plugins/vuetify.js as the source.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* refactor(browser): sum the browsePane toolbar offsets instead of enumerating them
Three toolbars can appear above the grid independently, so the padding
had eight cases and seven modifier classes spelling out every
combination. Because #browsePane is an ID (1,0,0) and a class is
(0,1,0), not one of them could win without !important — seven of this
file's nine, and seven of the repo's 76.
Each modifier now sets one custom property and the ID rule sums them.
Three classes instead of seven, no !important, and a fourth toolbar
would be one line rather than eight.
The 0px defaults deliberately live in the var() fallbacks, not on
#browsePane: custom properties obey specificity, so an ID-level default
would beat every class-level override.
Verified in the browser across all eight states — 134 / 174 / 166 / 206
/ 154 / 194 / 186 / 226 px — each identical to the class it replaces,
and the live search toggle lands browsePaneSearch and 166px.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
* refactor(style): one global stylesheet, zero unscoped style blocks
Four components carried an unscoped <style> block, each with an
eslint-disable above it, and two of them held byte-for-byte copies of
the same .coverPopup rules. An unscoped block in a component is a
global rule filed under whichever component happened to need it first,
and .settingsSubHeader proved it: six components across the browser and
reader drawers used a class defined in settings-drawer.vue, with
comments pointing back at it.
src/styles/global.scss holds all of it, imported once from main.js
after vuetify/styles. The three back-pointing comments now name the new
file. The !importants stay for now; they come out with the cascade
layers.
Verified in the browser: all nine rules present once in the document,
.settingsSubHeader still #808080 / bolder / 10px in the drawer, and
every .coverPopup declaration resolving to its old value — 70vh, 60vw,
4px, the same shadow.
Also adds the NEWS lines for the card hover border, the tagging chip
icon and the PWA app bar colour.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* fix(tagging): release the pooled Metron connections after each task Since mokkari 4.8.0 every session keeps one pooled TLS connection, and comicbox 5.2.0 releases it only when asked. Nothing in codex owns all of them: the scan builds one OnlineSession, the prompt applier builds its own per answer, and the explicit-id fetch lets comicbox build a third. So the release is process-wide, which is safe because every Metron path in codex runs on the one OnlineTagThread. run_session releases as the last statement of its finally, after the deferred applies — those run on this same thread and open their own sessions, so releasing before them would only force a reconnect. resolve_prompt releases in a finally around the apply, and not on the skip branch, which makes no request at all. The credential validator builds a raw mokkari Session per check in the web process and never closed it; now it does, on every path. A reconnect per answered prompt is the cost, and it is exactly what every request cost before mokkari 4.8.0. A scan still keeps its pool for the whole run. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(tagging): link an in-batch destination collision to its rival Two archives in one directory whose names differ only by suffix both convert to the same CBZ. Codex's own preflight passes them: it models the database, and no row or file holds that name yet. comicbox 5.2.0 finds it — its claim map refuses the second before any archive is opened — and now says which archive it collided with, as a typed DestinationOccupiedError carrying the occupant. That occupant is the one signal codex could not have produced itself, so the write-error record turns it into the same twin link #865 added for the database case. The destination is a fallback: a row may have appeared there between codex's check and the write. comicbox's own message stands — it already names both archives — and the batch still queues one errors-changed task rather than one per error. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(tagging): correct two docstrings comicbox 5.2.0 invalidated Both said comicbox refuses an occupied destination "only after opening the archive, merging the metadata and serializing it". 5.2.0 refuses before reading any metadata, and bulk_write sniffs every destination before the write pool starts. So the reason codex keeps its own check is not timing. It is the database: which comic holds the twin, so the error can link it, and the scheme-name twin a rename-on write would mint, which no filesystem check can predict. The one-message-per-path reasoning is unchanged and stays. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
…nent (#874) The browser table's cover cell and the admin custom-covers grid each hand-rolled the same v-menu, with byte-identical popup rules that #871 had already deduped into the global stylesheet. Both now use cover-popup.vue, which #872 landed. Both gain what neither had: hover-open, focus-open and keyboard access. An <img> is not focusable, so the only way to enlarge a cover was a mouse click. They also lose their hand-rolled @Mouseleave — Vuetify's own hover handling covers the content too, with open and close delays. The browser cell's popupOpen state goes with it; CoverPopup is keyed on row.pk, so a reused cell instance gets a fresh popup rather than one pinned to the previous row. The .coverPopup block leaves global.scss, which is the hand-off #871's WS8 left for this PR. Nothing references the class now: it was never global styling, only a component nobody had written yet. Two details the move forced: Each site's crop and radius go in as a style object rather than a class. CoverPopup's menu branch is a VMenu fragment, and a scoped parent class cannot reach a multi-root child — it would land without the data-v attribute and match nothing. CoverPopup gains a `loading` prop, defaulting to the lazy it already used. The table thumb is sized `width: auto`, so before its image loads it has zero width; a zero-area element never satisfies the browser's in-viewport check, and the load never starts. Caught in the browser: 0 of 17 covers loaded, against 17 of 17 on develop. It passes "eager". Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
* feat(style): declare the cascade-layer order A browser fixes a layer's position the first time it sees the name, and django-vite links imported chunks' CSS before the entry's own — the built manifest has three Vuetify chunks whose CSS opens with `@layer vuetify-components`. So production already establishes vuetify-components before vuetify-core, and any declaration shipped through main.js would arrive second. Inline in the Django template is the only placement that cannot be out of order and cannot 404. The order is the whole point. codex-base and codex-components sit after vuetify-components, so a codex rule beats component CSS with no !important, and before vuetify-utilities, so a `color` prop still wins. codex-components after codex-base lets a component's own rule beat the global `a` rule. codex-trumps is last, the named escape hatch; vuetify-final holds forced-colors accessibility fixes and is never overridden. Nothing is in any layer yet, so this commit changes no rendering. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(style): put the global stylesheet in the codex-base layer With the order declared, the snackbar carve-out can go. A snackbar's `color` prop paints .v-overlay__content through a `bg-*` utility class; an unlayered background-color beat that class no matter how the snackbar asked, which is why the rule had to exempt .v-snackbar__wrapper by name. In codex-base the utility wins on layer order, so the rule addresses every overlay and says what it means. That is the root cause of #845 rather than the per-instance patch it shipped with. The file's three !importants go with it: each existed to beat Vuetify's own layered component CSS, which codex-base now beats by position. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(style): the five shared partials declare their own layer A partial has to wrap itself. `@use` emits its CSS at the top of the consuming stylesheet, outside any @layer the consumer writes, so a consumer-side wrap would silently leave the shared rules unlayered — and unlayered CSS beats all layered CSS, so they would win over everything. `@use` inside @layer is a Sass error, so there is no third option. Verified against the built CSS: the layer survives @use and @forward, Vue's scoping lands the [data-v-*] attribute inside the block, and :deep() still compiles to a descendant selector within it. anchors.scss drops its two !importants with the move — they existed to beat Vuetify's own anchor styling, which codex-components now beats by position. That ends the anchor war: the four components whose own !important existed only to outrank anchors.scss drop theirs too. Their scoped blocks are unlayered, so they win outright. 76 !important declarations at the start of this plan, 62 now. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(style): layer the settings area's scoped blocks First of the six areas, smallest blast radius first. Three colour/background !importants go: two backgrounds that beat Vuetify's navigation-drawer CSS, and the :deep() list-item icon tint. All three lose to codex-components' position now, and the icon tint keeps winning over admin-menu's semantic colours the same way it did — that rule is more specific, and specificity still decides inside a layer. The `position: relative !important` stays: it is layout, not colour, and this audit is scoped to colour. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(style): layer the metadata area's scoped blocks No colour !importants to drop here — metadata-text's went with the anchor war in the partials commit. The two that remain set font-size and min-height, which this audit does not touch. Audit (b): edit-panel is one of only two files in the wrap set that combine a Vuetify spacing utility with scoped rules. Its `mt-2` checkbox is matched by no scoped margin rule, so nothing changes hands. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(style): layer the root components' scoped blocks Ten files. The audit turned up the first rule that genuinely has to beat a utility class: VEmptyState hands its action button `color: surface-variant`, which arrives as a bg-* class in vuetify-utilities, and utilities deliberately outrank codex-components. That is what codex-trumps exists for, so the rule moves there and drops its !important. Passing `color="primary"` to the empty state instead would let Vuetify derive on-primary as white and flip the button's black label — a look-at-it decision the theme plan deliberately deferred. The other !importants in this area set margins, position and opacity, which this audit does not touch. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(style): layer the admin area's scoped blocks Seven files, and the clearest payoff in the plan: admin-menu's notification icons needed `!important` purely to outrank settings-drawer's blanket icon tint, which itself needed `!important` to outrank Vuetify. Both rules are in codex-components now, so the extra class here wins on specificity alone and neither side carries !important. Three of them gone, and a comment that used to explain an arms race now explains ordinary CSS. library-table's `.disabled` goes with them. It is not in the wrap set — its !important fought the global anchor rule, which is layered below it now. Audit (b): tagging-status-table is the other file combining Vuetify spacing utilities with scoped rules. Its three `mr-1` elements are matched by no scoped margin rule. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(style): layer the reader area's scoped blocks Eight files, two colour !importants dropped: the arc select's selection text and the page-error empty state's icon and headline. Both fought Vuetify's own component CSS, which codex-components beats by position. Neither element carries a utility class — VEmptyState only colours its icon when a caller passes `color`, and nothing here does. The remaining !importants in this area set height and width on the PDF canvas and the page tracker, which this audit does not touch. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * feat(style): layer the browser area's scoped blocks Nine files, the largest blast radius, which is why it went last. Three colour !importants drop: the card control buttons' colour and hover colour, and the transparent v-btn overlay. Neither button sets a colour prop and VBtn is not in codex's defaults, so no utility class competes — all three only ever fought Vuetify's own component CSS. 76 !important declarations when this plan started, 50 now. The rest set geometry, size and opacity; this audit was colour only. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> * docs(design): record app-wide colour and the cascade DESIGN.md described the admin area's visual language. Colour and the cascade are not admin-only, so the scope line now says which sections apply where, and two new sections record what the code actually does. §9 Colour: the two ways to reach a theme colour and the four ways not to, with the contract test named as the thing that fails. §10 Cascade layers: the order, why it lives in the Django template and nowhere else, which layer each kind of stylesheet belongs in, the two Sass rules that make self-layering partials load-bearing, and that !important is no longer the tool for beating Vuetify. Every claim here was checked against the tree it describes. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
The SMTP host field required a dot, so `localhost`, `mailhog` and every Compose service name could not be typed into it even though the server accepts them -- `host` is a plain CharField with no format check. Drop the regex rule rather than loosening it; the server enforces length only, and a blank host is the documented "email is off" state. The change-password dialog read MIN_PASSWORD_LENGTH off the component while mapping only showChangePasswordDialog, so the comparison ran against `undefined` and the minimum-length check never fired. Same bug class as #867 in login-dialog.vue; fixed the same way. Had it fired the message would have read "Password must be undefined characters long". Its wording is unified with the other four password rules while here. The new vitest seeds a minimum of 6 so a missing mapping, a misspelled store key or a hardcoded 4 fails. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
A scan that cannot find a file infers a deletion, and the delete cascades the comic's bookmarks away. The reporter lost a publisher's read progress twice to a share unreachable for 41-50 minutes, far longer than the five-second second look covers. This is the state that makes retention expressible. Nothing reads or writes the column yet, so the change is inert: no behaviour moves until the hiding, stamping and revival PRs land on top of it. Two things here are easy to get silently wrong, and both are pinned by tests that fail when reverted: A subclass Meta's `indexes` REPLACES the base's rather than merging, so declaring the partial index on `WatchedPath.Meta` would hand it to Folder, FailedImport and CustomCover while silently skipping Comic -- the model that most needs it -- and would drop Comic's existing ACL index at the same time. Both concrete models declare their own. `BULK_UPDATE_COMIC_FIELDS` is derived from `Comic._meta.get_fields()`, so the column enrols itself and `bulk_update` would write a freshly constructed Comic's None over a live stamp on any unrelated re-import. It joins `metadata_imported_at` in the exclusion set. That also keeps it out of `BULK_CREATE_COMIC_FIELDS`, which is built from the same tuple. The index is partial on purpose. The visibility filter is `missing_since IS NULL`, which matches ~every row and which no index can serve; the reap's `missing_since < cutoff` matches ~none. Precedent is LibrarianStatus's partial index on its nullable datetime markers. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Three sites painted the orange themselves and set a literal black label, while every `bg-primary` element took Vuetify's derived `on-primary` -- which `genOnColors` picks as WHITE for this orange (verified by running hasLightForeground against #CC7B19). So the app had two different foregrounds on the same accent. Declare `on-primary: #000000` and point the three literal blacks at the token. Black everywhere: the three sites are unchanged, and six elevated/flat `bg-primary` buttons flip from white to black -- the SSO login button, the browser-settings Save, the table column-picker Save, the online-tag launcher, and the metadata edit-panel confirm and add-URL dialogs. Tonal buttons, chips and progress bars are unaffected; they use `text-primary` or carry no text. The three rules are replaced, not deleted: deleting them would leave the empty-state label on `on-surface-variant` and the list items on their inherited colour. empty.vue's comment about the derived white is now stale, so it records the real reason not to pass `color="primary"` there instead -- VEmptyState also hands `props.color` to the icon. Measured in the running app: `--v-theme-on-primary` resolves to 0,0,0; `.clearFilter` renders rgb(0,0,0) on rgb(204,123,25); the column-picker Save button renders rgb(0,0,0) where it was white. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Every validation number the admin client enforces is also enforced on the server, and the client re-typed them. That is how #867 happened -- a minimum-length rule reading a store key nobody had mapped, comparing against `undefined` and never firing -- and how it happened a second time in the change-password dialog, which this repo fixed last week. `codex/choices/limits.py` types each limit once and generates the client's copy, so a component cannot drift from the server and a misspelled key is undefined rather than stale. It deliberately imports nothing from Django or comicbox: `codex.settings` imports it, so it has to stay importable before the app registry is ready. A test asserts that. The password minimum is the proof case and the first consumer. Server enforcement is exclusively Django's `validate_password()` and no DRF serializer carries `min_length`, so OPTIONS metadata could never expose it -- a generated constant is the only transport that can. All five password sites now read `LIMITS.passwordMinLength` directly and the store key is deleted. The three field widths move here and `codex.models.base` imports them back, and the OIDC url literal is replaced by the constant of the same value, so no schema moves. The three widths Django owns are mirrored with pytest tripwires rather than definitions. The two dialog vitests change contract with it: seeding a store value no longer means anything, so they read the generated file and assert it is a real number -- which is what catches a typo'd key, the failure mode that started this. The remaining constants have no consumers yet; the bounds and length rules that use them land in the follow-ups. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
Three holes where a direct API call could save a value the app cannot survive. The client's rules were the only thing in the way, and a rule in a browser is not validation. A saved page size of 0 is a live 500: `get_browser_max_obj_per_page` honours "0" -- only an empty or unparseable value falls back -- and `ceil(count / 0)` raises ZeroDivisionError, so browsing stops working entirely until an admin fixes the flag from a UI they can no longer reach. A saved upload cap of -1 makes every custom-cover upload fail with "Upload exceeds -1 MB limit". Neither value is rejected anywhere today. Both live in a text column shared with every other flag, so a MinValueValidator cannot go on the field; `AdminFlagSerializer.validate` is the only place the bound can be expressed. It grows a per-key range map beside the existing per-key branch, and `validate` splits into two named helpers rather than growing a third arm inline. The SMTP port and timeout are not literally unbounded -- Django derives a range from the integer type and DRF lifts it into the field -- but the derived range is (0, 9223372036854775807), measured here, so port 0 and port 70000 both save while negatives are refused. Explicit serializer-level bounds, not model `validators=[...]`: validators deconstruct, so those would cost an AlterField migration for no schema change. On the client, `flag-card.vue` has no v-form and PATCHes on every keystroke, so a rule displays a message without stopping the write. The write itself is gated. Both client controls now read their bounds from the generated limits file rather than retyping them. Each test fails against the unbounded code. Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
`from_address` is the envelope sender of every outbound message and was
a CharField with client-only validation, so anything at all could be
saved through a direct API call and a malformed one failed at send time
instead of at save time. It becomes an EmailField on both email
serializers rather than on the model: the model type would be truthier,
but an AlterField migration buys nothing the serializer cannot enforce,
and `put()` is the only admin write path. `max_length` stays explicit
so it tracks the column rather than EmailField's own default of 254.
The test-send serializer's `host` and `subject_prefix` gain the widths
the save path already enforces and they carried not at all.
D6 suggested consuming DRF's OPTIONS metadata for the client's length
rules instead of new constants. It cannot work here: SimpleMetadata
emits field info only for `{PUT, POST} & allowed_methods` on views with
`get_serializer`, and in this codebase the detail routes are PATCH-only
while the email, OIDC and throttle singletons are plain AdminAPIViews
with no serializer hook. It would cover three of the six fields, miss
the SMTP host and the OIDC url entirely, cost a round trip per dialog,
and return snake_case under the JSON:API renderer. The generated
constants cover all six with no new alias -- Vuetify already ships
`$maxLength`.
No frontend rule enforced any length before this, on any field.
The SMTP host gets a rule again, the honest one: length only. The regex
removed last week refused `localhost` and every Compose service name;
the column width is the one thing the server actually checks there.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
D6 called this "a deletion, not an addition" -- stop scanning locally
and surface the server's error. But there was no server-field-error
plumbing to surface it into, on either of the two renderers, and
neither shape is the one a reader would expect.
Admin resource viewsets render through AdminJSONAPIRenderer, whose
`format_errors` is literally `{"errors": data}` -- so a 400 body is
`{"errors": {"username": [...]}}`, a dict rather than the JSON:API error
list, because DJA's own exception handler is not installed. The
interceptor only wraps when `errors` is an array, so it stays a plain
xior error, and `getErrors` scanned its known keys at the TOP level and
fell through to `errors = [data]`. SubmitFooter then rendered an object,
which Vue stringified as JSON.
Envelope endpoints emit `{"errors": [{status, title, detail}]}`, which
DOES wrap into an APIError -- and APIError has no `.response`, so the
same parser logged "Unable to parse error" and returned ["Unknown
error"]. The email tab showed no save errors at all.
Both shapes are pinned by a pytest written before any of this, and both
now normalize into one `{field: [messages]}` map that inputs bind
per-field. Keys arrive camelCased, matching the client's model keys.
`flag-card.setError` read `Reflect.get(this.formErrors[0], field)`
against a list of message strings, so it was always undefined and a
flag error could never render at all.
The client uniqueness scans stay as a fast pre-check rather than being
deleted -- #870 made them getter-based so they re-read after a
websocket reload -- but they were never the authority and are now
documented as such: `loadTable` reads only `body.results` and never
follows `next`, so past the first page of 200 they cannot see the
duplicate the server will reject. `library-create-update-inputs.vue`
never had an equality check at all, so it gains correct behaviour it
never had.
Co-authored-by: Claude Opus 5 <noreply@anthropic.com>
…996) A library change (e.g. Mylar importing into the series being read) reloads the open book through loadMtimes -> loadBooks. loadBooks stepped the prev/next page routes from params.page, but the reader route is /read/:pk and keeps the page in the query, so both routes came back false. The next-page link then resolved to /read/:pk with no page (the cover) and the arrow keys fell through to the book-change prompt, so a second press opened the next book. An arc switch mid-book did the same. loadBooks now steps from the store's page when reloading the open book (the query lags in cacheBook mode) and from the query's page otherwise. On a keyboard book change, the book-change flag pre-mounts the target book's pager, and its $route watcher called setActivePage before loadBooks swapped the books. setActivePage acts on the current book, so the book being left got the new book's page and bookmark: going forward reset it to page 0, going back marked an unread book finished. The pager now sets the page only once its book is the current book. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
ty 0.0.92 resolves Django foreign-key ``_id`` attributes, so 17 ``# ty: ignore[unresolved-attribute]`` comments went unused. Drop the ty part and keep each line's pyright ignore, which basedpyright still needs. It also rejects an in-method ``self.request: Request`` declaration as an incompatible override of Django's ``View.request``. AuthMixin, BookmarkFilterMixin and UserActiveMixin now declare it at class level, as APIView does; ``self.request`` still resolves to DRF's Request. SettingsReader.comic_id is now ``int | None`` (global rows have no comic). The reader's batch lookup filters on ``comic_id__in``, so a walrus guard narrows the key instead of an ignore. ``comic.pk = None`` is Django's copy idiom, but django-stubs types pk as never None, so that test line gets a narrow ty ignore. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…500 (#998) AdminUserViewSet's guards against deleting or deactivating the logged in user raised django.core.exceptions.ValidationError, which DRF's exception handler does not convert, so both escaped as a 500. In the admin UI, unticking your own Is Staff or Is Active and saving showed an unknown error instead of the reason. Delete is already hidden on your own row, but the API now answers it the same way. Raise DRF's ValidationError keyed by ``detail``. The JSON:API renderer wraps it as {"errors": {"detail": "..."}}, which the admin form shows as a plain sentence. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* Adapt to comicbox 5.5.0: MetronInfo series fields, key-only Metron, BCP 47
comicbox 5.5.0 writes MetronInfo v1.2, which carries a series' EndYear,
Status and Summary. The tag editor now writes them into MetronInfo like
Start Year, gated to that format, and the re-read imports them. The
database-only path that stood in for a format goes: POST
/admin/series-attrs, its serializer, the "saved in Codex only" hint, and
the ImportTask.series_attrs side channel that carried a tag write's or
online match's series values past the file. A ComicInfo-only write now
stores none of them, like every other MetronInfo-only field. Clearing one
empties the file but, as for Start Year, leaves the Series row.
MetronInfo refuses an EndYear on an Ongoing series or before StartYear,
and comicbox drops it with only a log line, so End Year gets a rule that
blocks the save with that reason.
Metron is API-token only in comicbox 5.5.0 / mokkari 5.0: the credential
dataclasses lost user/password and raise TypeError on them. Migration
0059 drops the stored login, every credential builder and "configured"
check is key-only, the deprecation UI goes, and backups stop carrying the
columns (old sidecars' are ignored). Telemetry keeps the
has_metron_credentials key, now from the key alone.
BCP 47 language tags like pt-BR now survive comicbox. They read as
"Portuguese (BR)" in the serializer, browser table and search index
instead of the raw tag plus a warning per row, and the editor maps the
name back to the tag.
The pyproject pin and lock wait for the comicbox 5.5.0 release.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* update deps
* Adapt to djangorestframework-jsonapi 9.0: own the admin error body
Codex renders admin resource viewsets with DJA's JSONRenderer but not
DJA's exception handler, so a 4xx reaches render_errors as DRF's raw
error body. Through 8.1, format_errors wrapped any body as
{"errors": data}. In 9.0 it passes a dict through bare and sorts a list
as error objects. A duplicate username came back as {"username": [...]},
and a list of message strings would crash the sort and turn the 400
into a 500.
AdminJSONAPIRenderer.render_errors now builds {"errors": data} itself,
so the wire shape the frontend and tests/test_admin_error_shapes.py pin
stays put whatever DJA's helper does. A new test covers the list shape.
Nothing else in 9.0 applies: IncludedResourcesFilter only concerns
DJA's own views and serializers and ?include=, which codex doesn't use,
and the polymorphic removals don't touch it.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
…1000) A new visitor's first page load and its API calls arrive together on one fresh session. Each looked the browser settings row up, missed, and inserted; the loser hit unique_settingsbrowser_session and answered 500. The browser row, its filters and its last route are now created in one atomic block, so a loser rolls back cleanly and no request sees a row without its related rows. On IntegrityError the lookups run again and the winner's row is used; with no row found the error is re-raised. The reader path already absorbed the same race through get_or_create. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(frontend): register Vuetify labs components Auto-import the labs components (VCheckboxGroup, VSplitBtn). A checkbox group passes its own color to its boxes even when unset, so give groups the primary color checkboxes already default to, and line groups up with a lone checkbox: drop the labs CSS's 16px label and 6px box indents and space inline boxes apart. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(choices): derive the tag write formats Add the write formats, titled by comicbox, to tagging-choices.json so the tag editor and the admin Tagging tab stop hand-writing the list. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(admin): require at least one default tag write format comicbox refuses a write with no formats, so an empty default made every file in an online-tagging session fail. Reject an empty list, a non-list and unknown format ids. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(tag-editor): formats as a checkbox group Replace the Formats chip select with checkboxes. The last checked format is disabled and the group says why, which replaces enforceMinFormat and its undeclared lastFormats. The hint sits beside the boxes on a wide toolbar and under them, with the toolbar wrapping, on a narrow one. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(admin): tagging Metadata Formats as a checkbox group The last checked format is disabled and explained, so auto-save can't write an empty list. A row already saved empty shows an error instead of auto-saving a default on open. The docs links stay in the field's hint. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(online-tag): sources as a checkbox group A source without an API key is disabled with a hint pointing to Admin > Tagging, and an empty selection explains why Search is disabled. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(browser): column picker categories as checkbox groups Each category is a group titled by its label. A click still goes through toggleColumn, so a new column lands in canonical position, not at the end. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(browser): Show Collection Levels as a checkbox group A click still sends one setSettings call with only the flag it flipped. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(admin): Defaults Show Collections as a checkbox group The locked top collection's reason becomes the group's hint. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat(frontend): register Vuetify labs components Auto-import the labs components (VCheckboxGroup, VSplitBtn). Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(admin): library Poll / Force Update as a split button Experiment. Poll is the button and Force Update its one menu item. Both confirmations sit outside the menu, which unmounts its content when it closes, so ConfirmDialog gains a controlled mode: bound to a modelValue it renders no activator and reports every close. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* fix(admin): folder list reports unlistable folders and its parent A folder the server can see but not list answered "Server Error" with a logged traceback; it is now a 400 "Permission denied" keyed to the path. Folders sort ignoring case, so dc sorts next to DC, and the listing names its parent folder (null at a filesystem root). ".." stays for the current picker. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * feat(admin): folder tree picker for Add Library Experiment. A VTreeview rooted at the server's default folder replaces the combobox picker: one click chooses and opens a folder, a Parent folder button climbs while keeping what was loaded, Enter on a typed path re-roots there, and Show Hidden Folders refetches. Folders an existing library covers are disabled and say so. Nothing is chosen on open, since the default folder is the newest library's parent. Past 200 children the rest collapse into a "type the path" leaf, which keeps a 5,000-folder directory opening in about 0.6s. The old picker stays until the go/no-go. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * fix(admin): Show Hidden Folders keeps the folder tree in place Toggling it re-rooted the tree at the chosen folder, collapsing everything above and beside it. It now refetches the root and each open folder with the new setting and keeps them open; a closed folder lists with the new setting when it is opened. Hiding never hides the way to the chosen folder, so unchecking it deep inside .claude keeps the open branch down to the choice. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * refactor(admin): remove the old folder picker The folder tree replaces it. With it go the store's folderPicker state, loadFolders and clearFolders, and the server's ".." entry: the tree climbs with parentFolder instead. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* docs(news): Keep a Changelog sections under gitmoji headings Every version's legacy `- Features` style list labels become `### ✨ Added` style headings in a fixed order: Breaking Changes, Added, Changed, Deprecated, Removed, Fixed, Security, Developer. About 25 legacy spellings fold into those; status-only versions (yanked, broken) move their note into the version title. cfg/eslint.config.base.js takes devenv's checkSiblingsOnly change so a section heading may repeat under each version but not within one. The v2.4.1 golden release body follows the new format; GitHub Releases already published keep their old bodies. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * restore some sections from old news. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* update deps
* chore(deps): bump eslint_d and eslint-plugin-yml
From the bun update that make update-devenv runs: eslint_d 15.1.0,
eslint-plugin-yml 3.10.0 and their lockfile transitive updates.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(devenv): update to 0.3.1
Raw make update-devenv output against devenv 0.3.1, plus uv lock for
the new docs group.
- mkdocs.yml, docs/requirements.txt and docs/style.css retired
- zensical.toml created from the devenv template
- .readthedocs.yaml builds through build.jobs with uv
- docs group: mkdocs + mkdocs-minify-plugin -> zensical~=0.1
- source-include: mkdocs and docs/style*.css entries out, zensical.toml in
- cfg/docs.mk runs zensical build --strict / zensical serve --open
- cfg/eslint.config.base.js ignores site/
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs: Zensical project settings, drop the mkdocs theme stylesheets
zensical.toml [project] takes site_name, site_author, site_description
and nav from the old mkdocs.yml, plus site_url for Read the Docs and the
codex/img/logo.svg theme logo. The style.{material,mkdocs,readthedocs}.css
variants only existed for docs/style.css, which devenv retired.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(build): sort source-include, drop entries docs/** covers
update-devenv appended docs/** and zensical.toml after the sorted list.
The explicit docs/ files are all inside docs/**; the sdist file list is
unchanged (2219 files, compared before and after).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* fix docs orange
* chore(deps): bun update transitive lint deps
From the bun update that make update-devenv runs: mdast-util-* and
vue-eslint-parser patch releases. package.json is unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(devenv): update to 0.3.2
Raw make update-devenv output against devenv 0.3.2: the docs palette
primary moves from deep purple to indigo (accent orange stays), and
zensical.toml enables admonition, pymdownx.details, pymdownx.superfences
and pymdownx.quotes callouts.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* chore(devenv): update to 0.3.3
Raw make update-devenv output against devenv 0.3.3: Markdown lint
accepts GitHub callouts (no-missing-label-refs allowLabels for !NOTE,
!TIP, !IMPORTANT, !WARNING and !CAUTION). The bun update and uv lock
changed nothing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs: GitHub-style warning callouts for the deprecation and caveat notes
The Docker Hub move note, the three config deprecation notes and the
read-only /comics mount note become > [!WARNING] callouts, which GitHub
renders natively and Zensical renders as admonitions through
pymdownx.quotes callouts. The bare > line after each marker keeps
Prettier's proseWrap from joining the marker onto the text.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs: note that a correct proxy needs no CSRF trusted origins
The CSRF 403 section now ends with a callout: in general a correctly
configured reverse proxy shouldn't need csrf_trusted_origins or
CODEX_CSRF_TRUSTED_ORIGINS.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
* docs: lead the CSRF 403 section with its note
The note that a correctly configured reverse proxy needs no CSRF
trusted origins moves from the end of the section to the top, and the
first paragraph drops the clause that said the same thing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---------
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Wrap every 💥 Breaking Changes section in a > [!WARNING] callout and give each yanked or broken release a > [!CAUTION] body naming the release to use instead, via devenv's news_normalize.py. Also point v2.3.1 at the 💥 Breaking Changes sections instead of the long-gone Upgrading notes, and normalize two sub-group labels in v1.2.4. Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* chore(deps): comicbox 5.6.0 comicbox 5.6.0 changes only the filename format: filenames now read the month, day, volume count and publisher, renames write the full date and volume count but never the publisher, and `(Digital)` / `(DC)` alone no longer parse as the title. comicfn2dict 1.1 and comicbox-pdffile 1.1.2 come along. codex needs no code change. None of the new filename keys is a new top-level comicbox field, so the derived delete_keys list is unchanged and `remainders` is still dropped before a rename. The rename preview goes through predict_filename, so it shows the new names as is. Checked on metadata-less CBZs through COMICBOX_CONFIG: publisher, full date and original_format are imported, and the predicted names match comicbox's new format. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * chore(ty): rename unresolved-import ignores to import-not-found ty 0.0.99, already locked on develop, renamed the rule. The old name now raises renamed-rule-in-ignore-comment, which fails `make ty`. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> * comicbox 5.6.1. --------- Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The tag editor's publisher, imprint, series, volume, credit, story arc, universe and tag inputs only offered the opened group's own values, so correcting a name meant retyping the library's spelling exactly; a typo silently created a new row. New admin endpoint GET /api/v4/admin/tag-names/<field>?q= returns at most 100 existing names containing q, prefix matches first, with a `complete` flag. Imprint, series and volume narrow to the form's parent names. A TagNameCombobox wrapper fetches only when its menu opens or the user types with it open (debounced, aborting the stale request). Results cache per field, scope and query in the metadata store until the dialog closes, and a complete result for a shorter prefix answers longer queries without another request. Free text is still accepted. Claude-Session: https://claude.ai/code/session_01UgiA5Rda3grk8WEu1YhY4S Co-authored-by: Claude <noreply@anthropic.com>
* feat(tag-editor): suggest existing names as you type The tag editor's publisher, imprint, series, volume, credit, story arc, universe and tag inputs only offered the opened group's own values, so correcting a name meant retyping the library's spelling exactly; a typo silently created a new row. New admin endpoint GET /api/v4/admin/tag-names/<field>?q= returns at most 100 existing names containing q, prefix matches first, with a `complete` flag. Imprint, series and volume narrow to the form's parent names. A TagNameCombobox wrapper fetches only when its menu opens or the user types with it open (debounced, aborting the stale request). Results cache per field, scope and query in the metadata store until the dialog closes, and a complete result for a shorter prefix answers longer queries without another request. Free text is still accepted. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UgiA5Rda3grk8WEu1YhY4S * chore(release): v2.6.0 - Tag Editor Suggestions Bump to 2.6.0 from the unreleased 2.5.8 and retitle its NEWS section, leading with the tag editor's name suggestions. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01UgiA5Rda3grk8WEu1YhY4S --------- Co-authored-by: Claude <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
✨ Added
volumes, credits, story arcs, universes and tags as you type, so a correction
can pick the spelling already in use instead of creating a near-duplicate.
Imprint, series and volume suggestions narrow to the publisher, imprint and
series chosen above them. New names can still be typed in.
Add Libraryserver folder picker component.♻️ Changed
(1950-11-01), and the volume count(comicbox 5.6.0).
🐛 Fixed
(Digital)or(DC)alone no longer read as the title (comicbox 5.6.0).Error".