DOCKER_HOST only decides which daemon creates the container; the address the
test dials comes from POSTGRES_HOST, which defaults to the job container's own
loopback. Setting one without the other still times out.
The runner has docker, so the database container started; the job just could
not reach it. pytest-databases takes the address from DOCKER_HOST, and unset
means 127.0.0.1 -- the job container's loopback, not the namespace the sibling
published on. A dind service makes it resolve to a host that answers.
CI has no nix shell, and ruff was only ever a nix package, so both lint steps
died with "Failed to spawn: ruff". Pinned exactly: the formatter deciding what
is correct must be one version, or CI rejects what the dev shell produced.
Only src/ was covered, which is why it stayed clean while tests/ drifted to 38
findings. migrations/ stays out: its alembic template emits unused imports, so
every generated revision would fail the gate.
Mostly automatic: empty f-strings, unused imports, formatting. The manual half
was duplicated import blocks stranded mid-file and `== None` assertions, which
become `is None` here because they compare plain attributes -- unlike the
identical rule in filters/book.py, where they build SQL.
Two unused variables are the tests never checking the disk in either
delete_files direction. Left in place under a noqa so the gap stays visible.
svelte-check is clean, so it gates like the rest. eslint still reports without
failing, now for two findings rather than 87: both are the unsanitized book
description, written up in TODO.md.
asChild is a bits-ui v1 prop the icon picker no longer needs -- it already
passes a child snippet. The drop zone set both webkitdirectory and a bare
directory attribute, which no browser implements. The library view seeds its
collection state once by design, so it says so with untrack.
Dead imports and locals removed, each blocks keyed, `any` narrowed to unknown.
Two context setters kept their calls and lost only the unused binding; the
settings redirect no longer awaits a parent whose data it discards. A leading
underscore now marks a binding that only holds a position.
no-unused-expressions is off for Svelte files: a bare `book;` in an $effect
declares a dependency. The URL and Set instances it flagged are local
temporaries, marked individually. The one real finding was visibleApiKeys,
genuine reactive state cloned on every mutation to force an update; it is a
SvelteSet now.
The RequestHandler import sat commented out while all four handlers were
annotated with it, so every destructured parameter was an implicit any --
24 of the 30 svelte-check errors. deleteFn returned `{}`, which void is not
assignable to.
no-useless-assignment joined the recommended set and does not understand runes,
reading every `$bindable()` prop default as a dead store; off for Svelte files.
Carries @eslint/js, @eslint/compat, globals and @types/node with it.
SvelteKit 2.53 to 2.70 renamed the remote form enhance callback's `form` to
`element`, which would have thrown on reset in the signup, device, library and
cover forms.
Every push runs ruff, prettier, pytest and svelte-check; a tagged release runs the
blocking half again before it publishes an image. eslint and svelte-check report
without failing, since 87 and 30 findings predate the workflow.
Applies ruff format, ruff check --fix and prettier, so CI can gate on all three.
The surviving unused imports were all re-exports in __init__.py, now covered by a
per-file ignore; the pdf.js viewer and the generated openapi types join the
vendored code that eslint and prettier already skip.
Tagging v* builds both images on Gitea Actions and pushes them to the instance
registry, then smoke-tests the published stack. The frontend read its backend
URL through import.meta.env, which Vite resolves at build time, so an image
could only point at whatever the build host had; it now reads it at runtime.
Reads metadata.db and copies the books into a library — from a zip uploaded on
the library settings page, or from a path with `litestar calibre-import`. The
source is never touched, and re-running only picks up what is new.
Also names the formats mimetypes does not know: a Calibre library is full of
MOBI and AZW3, and a null content type used to fail the book endpoint.
A workbench with the folded records in a rail, the survivor's fields live and
editable beside them, and per-field actions chosen by what kind of field it is.
Fields the records agree on stay out of the way. Reachable from a duplicate
group and from the selection toolbar, which is the only way to merge a pair the
detector never proposed.
Files move on disk before any row moves, then everything repoints in one
transaction and the folded records are deleted. Progress keeps whichever is
furthest, shelves and tags union, and identifiers move only under a name the
survivor lacks. Metadata is left alone unless the caller resolves it, since
choosing between two titles is a judgement this endpoint cannot make. Nothing
is removed from disk.
advanced_alchemy annotates the `ids` query parameter as list[str] whatever the
config says, so a bigint primary key was compared against strings and Postgres
refused. Nothing had called ?ids= until now.
"Fluent Python, 2nd Edition" is one book with a field for the edition. Left in
the title it also splits the library, since the second edition never looks like
the first. Only a numbered statement is moved, so "Catch 22" keeps its number
and "Global Edition" — which has nowhere to go in an integer column — stays put.
Identifiers are a collection, but the whole dict was replaced per format, so
the last file to report won outright — an EPUB declaring an ASIN, a Google id
and a Calibre id kept none of them once a PDF contributed one ISBN. Accumulate
instead, letting a declared identifier outrank one scraped off a page.
Extractors wrote whatever the file said, so one person held several rows:
"Sam Newman" beside "Newman, Sam;" beside "Sam Newman.epub", the last because
the upload path never stripped the file extension. Tidy on write, in the
validator and in the uniqueness lookup alike, and merge the rows that collide.
"Building Microservices, 2E" never matched "Building Microservices". Key the
strip on the trailing "e" so 2E, 5e and 3 Ed are caught, while a bare number
leaves "Catch 22" and "Blade Runner 2049" alone. Stored keys are recomputed.
Adds the per-library review screen, a muted note in the upload tray for a book
that may already be held, and the remote functions behind them. The tray note
is deliberately weaker than the file-level one, which means something stronger.
Match on a shared identifier, or on a normalised title credited to a shared
author, and report candidates rather than refusing anything — a metadata match
is a guess, and a second edition is not a mistake. Adds a library-wide review
pass, dismissals, and renames the fingerprint pre-flight to duplicate-files.
Derive normalized_title, normalized_name and normalized_value with @validates
so no write can bypass them, and add the table recording pairs a reader has
said are not the same book.
DC:identifier was validated verbatim, so hyphenated and urn:isbn: forms never
reached the checksum and every non-ISBN identifier was discarded. Normalise
first, and name whatever survives.
Reduce a title, an author and an identifier to a single comparison key, so
two copies of one book can be recognised by equality rather than by a
similarity score.
Storing the same bytes twice splits reading progress and shelves across two
records that can never converge. The skipped row still names the book that
holds the file; allow_duplicates stays on the API for a wrong verdict.
The hash lives in the database and the file does not, so a file deleted behind
the app's back went on refusing its own replacement. Matches are now checked
against disk, and re-adding a book's own missing file writes it back into the
row that already describes it.
Incoming files are matched against what is already stored, keyed on the
KOReader hash and the file size. Bulk uploads skip and report them, deliberate
creates are refused with a 409, the consume directory parks them aside, and
allow_duplicates overrides all three.
Also: books whose metadata generates a path another book already owns are moved
aside, so a forced copy cannot overwrite the original's files.
The filepath extractor was merged on the right, so a parent folder's name
overrode the title inside the file. Grouping also split a book's formats when
its own folder was the upload root.
file.path is relative to book.path, so the archive matched nothing on disk and
came out empty. Also namespace entries per book to stop filename collisions, and
stream the zip instead of buffering it whole.
patrick