The row-style decorators were unit-tested as pure functions and the populated test apps do carry a downloading queue entry, but text snapshots discard colour, so nothing proved the draw functions actually call the decorator with the live downloads list. Passing an empty slice instead left every test green.
Render the lidarr tracks table, the sonarr episodes table and the radarr movies table into a Buffer and assert the probe row's foreground: downloading_style with the queue populated, missing_style once the queue is emptied. The probe is the second row so the selection highlight never masks its style.
DownloadStatus was declared four times, byte for byte identically, in
radarr_models, sonarr_models, lidarr_models, and readarr_models: the same
variants, the same derive list, and the same serde, strum, and display_style
attributes. Every servarr's queue endpoint reports the same status
vocabulary, so there is nothing servarr specific about the type.
Move it into servarr_models, preserving that file's alphabetical ordering
(between DiskSpace and IndexerSettings), and drop the four copies. The
DownloadRecord types keep their per-servarr shapes and now reference the
shared enum.
The two enum tests, test_download_status_display and
test_download_status_to_display_str, were likewise duplicated in the four
per-servarr model test files. Keep one copy of each in servarr_models_tests
alongside the other shared enum display tests and delete the rest. The
per-servarr DownloadRecord fallback deserialization tests stay where they
are since they exercise each servarr's own record type.
Behavior is unchanged. Every remaining edit is an import path update.
The shared download record held id and episode id both equal to 1, and
the shared episode held id and episode file id both equal to 1, so
reading the wrong one of those fields could not fail any test that uses
them. Move the download record to id 9 and the episode to id 11, keeping
the download record's episode id equal to the episode's id so the
downloading episode still matches, and move the episode file to id 7
together with the episode's episode file id so the season view still
finds its quality and size. Series id, season number, and episode number
stay at 1. The handler tree keeps its own copy of these fixtures; update
it identically, along with the JSON twins in the network tests and the
seven handler assertions that named the old ids. The two get episodes
tests that asserted the sorted table now spell out the sorted order,
since the fixture episode no longer sorts first by id.
The shared track held id, artist id, track file id, and album id all
equal to 1, and the shared history item held id, album id, artist id,
and track id all equal to 1, so reading the wrong one of those fields
could not fail any test that uses them. Move the track to id 5, artist
id 3, and track file id 7 (with the shared track file moved to id 7 so
the album view still finds its quality and audio info), and move the
history item to id 13, artist id 3, and track id 5. Both album ids stay
at 1 so the fully populated album view still shows its track as
downloading. Update the seven assertions that named the old ids.
Movie.status and AddMovieSearchResult.status were bare Strings compared
against the literal "released" in the row-style decorator and in the
movie details style function. They now carry a MovieStatus enum with the
five values Radarr actually sends (tba, announced, inCinemas, released,
deleted), shaped like MinimumAvailability and DownloadStatus, and guarded
with from_json_or_default so an unrecognised status degrades to Tba
instead of failing the whole fetch.
Radarr never sends "Downloaded" as a movie status; its web UI derives
that label from released plus hasFile. Every fixture that carried
"Downloaded" now holds released, the value it was standing in for, so
none of them degrade to the default through the guard.
The status is never rendered as text in the Radarr tree, so no snapshot
changes; the CLI JSON output is unchanged since the enum serializes to
the same camelCase strings. The style_from_download_status tests now
exercise the unreleased arm, which previously had no case, in place of a
duplicated unmonitored-missing row.
A page is 20 rows and the index wrapped modulo one less than the row
count, so the last row was unreachable by paging, 21-row and 5-row tables
did nothing on page down from the first row, and 4-row tables cycled
0, 2, 1 and never reached row 3. Page down now clamps to the last row and
page up clamps to the first.
With no division left, a single-row table is safe, so the early return
only guards the empty table again, matching scroll_to_top and
scroll_to_bottom.
Radarr sends and accepts the credit type as "type" in both directions,
but the field was only renamed for deserialization, so serializing a
Credit emitted "creditType" and a to_value round trip could not be read
back. Credit is only deserialized in production, so this affects tests
and any future serialization of the record, not the live API traffic.
Radarr was the only tree still holding DownloadRecord.status as a bare
String and comparing it against literals. It now carries its own
DownloadStatus enum, shaped exactly like the Sonarr, Lidarr and Readarr
copies, and is guarded with from_json_or_default so an unrecognised
status degrades to Unknown instead of failing the whole queue fetch.
The row-style decorator and get_movie_status compare against the enum
variants. The status cell is never rendered in the Radarr tree, so no
snapshot changes; the CLI JSON output is unchanged for every real value
since the enum serializes to the same camelCase strings.
The shared download record held id, album id, and artist id all equal to
1, so reading the wrong one of the three could not fail any test that
uses the fixture. Move artist id to 3 and id to 9, keeping album id at 1
so the fully populated album view still shows its track as downloading,
and update the three handler assertions that named the old id.
decorate_with_row_style had no unit tests, so every one of its seven
return sites could be swapped for another style without a failure. Add
one test per branch, including the ordering cases where the download
match precedes the monitored guard and the monitored guard precedes the
air date check, using a local download fixture with distinct episode,
other-episode, and download ids.
The shared Book fixture left statistics as None, so every author details
snapshot rendered the missing-statistics N/A cells and the populated
branch of the Files and Size columns was only exercised by one test.
Give the fixture distinct, non-default counts and a size that formats to
a value different from the dedicated statistics test.
A Servarr build can emit an enum value this app has no name for, either
as a string it does not recognise or as a bare integer for a value the
build itself cannot name. Until now any such value in a response failed
the whole payload, so one unexpected artist status, download status,
task name, series type, minimum availability, credit type, or security
setting emptied the entire view with "Failed to parse response!".
Every enum-typed field on a struct the app only ever deserializes and
displays now uses from_json_or_default, so an unrecognised value
degrades to that enum's default and the rest of the record is kept.
Known values still map exactly as before and serialization is untouched.
Fields on request bodies (Add*Body, Add*Options, Edit*Params, the root
folder bodies) are deliberately left alone: those structs are only ever
serialized, so the attribute would be meaningless there. No candidate
struct is round-tripped: every edit and toggle in the app re-fetches
the raw JSON value and patches it by key, so the fallback can never be
written back to a server as a real setting.
The one Option field, the security config's authentication_required,
also carries serde(default), because deserialize_with on an Option turns
a missing key into a hard error. An unknown value there yields None.
Verified by mutation. Removing the attribute from one field per tree
fails exactly that field's test and nothing else. Dropping default from
the Option field fails only the missing-key test with "missing field
authenticationRequired"; dropping deserialize_with from it fails only
the unknown-value test. All 21 new fallback tests fail against the
unguarded code with "invalid type: integer 2, expected string or map".
Lidarr queues whole albums rather than individual tracks, so a track row
is styled from the queue entry whose albumId matches the track's albumId,
mirroring the Sonarr episode row logic. Tracks that already have a file
keep the downloaded style regardless of the queue.
The season, album and book tables rendered 0/0 and 0.00 GB whenever the
API omitted the statistics object for a row, presenting fabricated counts
as real. The row colour already treated that case as indeterminate; the
cell text now does too, rendering N/A for both cells, matching what these
tables already do for a missing release date.
Sonarr's season row is restructured from a destructure-with-default to the
same Option-aware map_or_else shape lidarr and readarr use.
The readarr test fixture book has no statistics, so its existing author
details snapshots change from the fabricated zeros to N/A on exactly that
row.
handler_proptest.rs was committed without ever being declared as a module, and it imported a list_size strategy that did not exist, so it never compiled. The proptest recipe filtered on a test name substring that matched no test.
Declare the module, add the two strategies it needs, and pass both property test filters to libtest so the recipe runs the ui and handler property tests.
page_up and page_down wrap the selection with a modulus of items_len - 1,
which is a remainder by zero when the table holds exactly one row. Ctrl-U
or Ctrl-D on such a table panicked the TUI. Both now return early for
tables of one row or fewer, mirroring the existing empty-table guard;
tables of two or more rows keep their existing paging arithmetic.
The add author confirmation popup previously left the space above the
root folder field blank. It now renders the selected search result's
overview there, in a fixed six line area matching the Radarr add movie
prompt, so the user can confirm they picked the right author before
adding. An author without an overview keeps the same layout.
The fully populated search result fixture gains a long overview so the
snapshot proves the paragraph is clipped to six lines.
A single history record whose event type the Servarr build has no name
for made the entire history response fail to decode, so the view died
with "Failed to parse response!" and showed nothing at all.
The event type is stored as an integer and is only serialized as a
camelCase string when the build can name it. For values it cannot name,
it is emitted as a bare JSON number instead, which is a type mismatch
rather than merely an unknown name. Serde's other attribute does not
help: it is rejected outright unless it sits on the last variant, and
even there it only catches unrecognised strings, leaving a number to
fail exactly as before.
from_json_or_default buffers the field and falls back to the type's
default for any input the enum does not recognise, whether that is an
unknown string, a number, or any other JSON value. It is applied to the
event type of all four history item types. Known values still map
exactly as they did, and serialization is untouched.
Verified against a live Readarr carrying such a record: the history
listing failed to decode before the change, and afterwards lists every
record, rendering the unnamed one as Unknown.
Verified by mutation. Seven mutations all fail the new suite: dropping
the fallback, discarding the decoded value, removing the attribute from
each of the four history items, and renaming a variant on serialize
only. The five per-tree and serialize-only mutations were each invisible
to every pre-existing test.
Toggling book monitoring failed with an HTTP 500, "Value cannot be null.
(Parameter 'source')", every time, in both the CLI and the TUI.
GET /book/{id} omits the editions field entirely, but PUT /book/{id}
requires it, so round-tripping Readarr's own response back as an update
was always rejected. Editions are served separately from
GET /edition?bookId=, so fetch them and attach them to the body before
the PUT.
The snapshot harness renders to a plain string, so colour and text modifiers never reach the assertion. The track row's downloaded and missing arms were therefore invisible to the suite: swapping either arm for the other, or inverting the has_file condition, all left the full run green.
Move the branch into decorate_track_row_with_style, mirroring the row decorators already extracted in the other library detail views, and unit test both arms directly against Row equality. All three mutations now fail the suite.
A monitored season rendered in the fully downloaded colour whenever Sonarr returned no statistics for it, and whenever it genuinely had zero known episodes. Both cases compared zero against zero and short circuited to the downloaded style, claiming the user had every episode of a season that was only announced, or that had never been scanned.
Seasons whose statistics are absent now render as indeterminate, matching the series row and the equivalent album and book rows in the other trees, and the downloaded style now requires a positive episode count.
Row styling is invisible to the snapshot harness, because render_to_string_with_app returns a plain String and discards colour and modifiers. The season row style branch was therefore entirely uncovered, and any arm could be swapped for another without a single test noticing.
decorate_season_row_with_style now holds that branch, matching decorate_series_row_with_style in the library view. The extraction is behaviour preserving: the branch is token identical to the one it replaces, the statistics fallback resolves the same way, and no snapshot changed. The season row deliberately has no positive episode count guard, so a season without statistics falls back to zero of zero episodes and reads as downloaded; a test now pins that.
Verified by mutation. Against the previous tree all seven mutations left the suite green: replacing each of the unmonitored, downloaded, unreleased and both missing arms, flipping the episode file count comparison, and inverting the monitored check. Against the new tests every one of those seven fails.
Row styling is invisible to the snapshot harness, because render_to_string_with_app returns a plain String and discards colour and modifiers. The album row style branch was therefore entirely uncovered, and any arm could be swapped for another without a single test noticing.
decorate_album_row_with_style now holds that branch, matching decorate_artist_row_with_style in the library view and decorate_book_row_with_style in the Readarr author details view. The extraction is behaviour preserving: the branch is token identical to the one it replaces, and no snapshot changed.
Verified by mutation. Against the previous tree all eight mutations left the suite green: replacing each of the unmonitored, downloaded, unreleased, both missing and indeterminate arms, flipping the track file count comparison, and inverting the monitored check. Against the new tests every one of those eight fails.
The collapsed one-line overview on the Lidarr artist, Sonarr series, and Readarr author details views replaced every line break and tab with an empty string. Words on either side of a paragraph break fused together, rendering as "20th century.She was born in Ohio".
Collapse each run of whitespace to a single space instead, and trim the result. Handling a whole run at a time keeps a Windows line ending, an indented paragraph, or a space that already precedes a line break from producing two spaces.
The overview popups are unaffected: they receive the raw text and keep their paragraph breaks.
The Radarr system details tests prefix their loading cases with renders_. This one was copied verbatim from the other trees and did not match. Rename it and its snapshot to fit, leaving the sibling trees alone.
draw_logs read the level from the log line's Display output, which drops the first offset characters. Once a log line is scrolled, the field at index 1 is no longer the level: past the timestamp it yields the logger name and colours the row wrong, and past the final delimiter the index is out of bounds and panics, taking the TUI down.
Parse log.text instead, matching what draw_logs_popup already does. The displayed text is unchanged.
The movie overview was the only overview popup drawn at the small size;
the sonarr, lidarr and readarr overview popups all use the large size.
The small size wrapped longer overviews sooner and gave the text roughly
a quarter of the space the other trees allow.
Draw it at the large size to match them. The wider popup now covers the
collection details behind it, as the other overview popups do.
The updates popup only checked whether its text was empty, so during a
refresh it kept rendering the previously fetched text. Opening the system
updates view dispatches a request for the update notes, so the popup could
sit on stale content until the response landed.
Guard on the loading flag as well, matching the sonarr, lidarr and readarr
updates popups, and add the loading-state test those trees have so both
halves of the condition are exercised. The pre-existing empty-updates test
never reaches the second half on its own.
The artist, album and season statistics fixtures set the file count equal to the total count, so the columns rendering the pair showed the same number twice and swapping the two fields changed nothing. The JSON fixtures move with them so the round trip assertions still hold.
The delete params tests set the delete files flag and the list exclusion flag both true, so swapping the two fields in the params builder produced identical params and no test noticed. The tests that assert a reset still set both true, since that is what proves both flags are cleared.
IndexersUi dispatches to IndexerSettingsUi, but no test rendered that route through the parent, so deleting the dispatch arm in radarr, sonarr and lidarr left the whole suite green.
Radarr was the only tree without a test rendering draw_updates_popup with no update text, leaving its loading branch unreachable from the suite. Forcing the guard true went undetected.
draw_updates_popup shows the loading block when update text is already present but a refresh is still in flight. Both existing tests fed empty update text, so they only exercised the first half of the guard and dropping the is_loading conjunct went undetected.
test_handler_delegation! asserted only that the current route equalled the
base block after Esc. When the delegated block was itself the base block the
stack read [base, base], so that assertion held whether or not the sub-handler
popped, and the case asserted nothing.
Capture the navigation stack depth before delegating and assert it drops by
exactly one, which also catches a handler that pops and then re-pushes. Depth
is read through a test-only accessor on App; the navigation_stack field stays
private.
Main tab blocks clear errors on Esc rather than popping, by design, so they
cannot be probed this way at all. Drop them from the delegation value lists;
the remaining values in each list already exercise the same dispatch arm.
ReadarrData::test_default_fully_populated seeded both profile maps with the same id and the same name, and gave the author fixture the same id for both profiles. The library table renders the two profiles in adjacent columns, so both read Standard and swapping the maps produced identical output.
The maps now use the disjoint ids and the vocabulary the root folder and add author tests already use, so the columns read Standard and Comprehensive.
Verified by swapping the two map lookups where the authors table is built. That swap survived the whole suite before this change and is now caught by sixteen snapshots.
Both tests seeded a single default task and asserted the extracted name equalled the default task name, so a body that ignored the table entirely and returned the default satisfied them. Radarr and Readarr already seed distinct named tasks and select one; these two now follow that shape.
Verified by replacing both bodies with a bare default task name. The Lidarr test could not have failed before, since it compared the default against itself, and the run confirms it is the only Lidarr test that catches the change. The Sonarr body was already caught by the two start task prompt tests, so there the dedicated test is strengthened rather than newly load bearing.
The handler fixtures for a collection movie and a searched movie both carried tmdb_id 1234 and the title Test, so the two branches of build_add_movie_body were value equivalent. The builder test also populated only the table its branch needed, so swapping the branches was caught by an unwrap panic on the empty table rather than by any assertion on the values.
The searched movie fixture now has its own identity, and the builder test seeds both tables so the branch is chosen against real competing data. Swapping the branches now fails with a value diff naming the wrong tmdb_id and title.
Only the handler copy of the fixture is changed. The network copy deliberately shares tmdb_id 1234 with the movie fixture, which is what drives the already in library marker in the add movie snapshots, so changing it there would have silently removed that marker.
Two of the three rows carried the quality EPUB, so sorting by quality left them tied and any reordering confined to that pair was invisible. The rows now use EPUB, AZW3 and MOBI, matching the vocabulary the history UI fixture already uses.
This one is defensive: no natural mutation of the quality comparator flipped from surviving to failing, because two distinct values already separated the cross-field cases. The tie is removed on principle rather than to close a demonstrated hole.
The Radarr, Sonarr and Lidarr tests for editing all indexer settings expected every numeric field to be 1 and returned every previous field as 2. Because the fields were mutually equal, transposing any two of availability_delay, maximum_size, minimum_age, retention and rss_sync_interval in the CLI parameter builder produced an identical request and no test could tell. Radarr additionally expected both boolean flags to hold the same value, hiding a transposition of those two.
Readarr already used distinct values and was the model followed here.
Verified by transposing minimum_age and retention in all four CLI parameter builders at once: before this change only the two Readarr tests failed, while Radarr, Sonarr and Lidarr stayed green. After it, the same mutation is caught by five tests across the three trees.
Sonarr and Lidarr each keep their own copy of the shared fixture, both with minimum_age and retention set to 1, so transposing those two fields stayed invisible in either tree. Radarr built its fixture from IndexerSettings::default(), leaving availability_delay, maximum_size, minimum_age and retention all 0: four mutually transposable fields, none of which could distinguish a threaded value from a hardcoded 0.
The no-op guard test seeded IndexerSettings::default() as its sentinel, which cannot show whether the guard held or the field was simply never written. It now seeds a distinct stale_indexer_settings(), mirroring the Readarr convention.
Verified by transposing minimum_age and retention in the Lidarr indexer settings UI, now caught by its snapshot, and by inverting the Radarr no-op guard, now caught by both GetAllIndexerSettings tests.
minimum_age and retention were both 1, so transposing them in any code that reads both rendered identically and no test could tell the difference. id was also 1, which cannot distinguish a correctly threaded value from a hardcode.
Verified by transposing minimum_age and retention in the Sonarr indexer settings UI: that mutation survived the full suite before this change and is now caught by the settings snapshot.
Readarr handler tests never asserted which block the search and filter keys push, so transposing searching_block and filtering_block in a TableHandlingConfig went undetected. History already had both tests; add the five missing search key tests and the three missing filter key tests, covering the authors, books, author history, editions and book history tables. The books and editions tables configure no filtering_block, so they get no filter test.
No test moved a table or list selection off index 0 either, which made current_selection() and items[0] indistinguishable. Select a non-zero index in the tests covering every handler that reads a selected row, adding a decoy entry at index 0 so existing expected values stay meaningful. The profile, monitor and root folder lists inside the add author, edit author and add root folder prompts are a separate gap that a table-only sweep misses, so those are covered too, as are the book files, editions and releases tables held on the book details modal.
Several fixtures also asserted against a default value, so the expected result was indistinguishable from an unset field. The add author, edit author and add root folder bodies now expect non-default monitor options, and the quality and metadata profile ids are drawn from disjoint id pairs so the two lookups cannot be transposed.
The search and filter blocks a table configures were asserted nowhere, so
transposing them was invisible. Selections were also left at index 0 in
every test, which made current_selection() and items[0] indistinguishable,
including for the profile lists inside the edit and add prompts.
Sonarr handler tests never asserted which block the search and filter keys
push, so transposing searching_block and filtering_block in a
TableHandlingConfig went undetected. Add a search key test for each of the
six searchable tables and a filter key test for each of the four filterable
ones, covering library, history, series details and season details.
No test moved a table or list selection off index 0 either, which made
current_selection() and items[0] indistinguishable. Select a non-zero index
in the tests covering every handler that reads a selected row, adding a
decoy entry at index 0 so existing expected values stay meaningful. The
modal list selections in the add and edit series prompts are a separate
gap that a table-only sweep misses, so those are covered too.
Several fixtures also asserted against a default value, so the expected
result was indistinguishable from an unset field. The system tasks now use
two distinct non-default task names, the season monitoring and manual
search expectations use distinct non-zero ids, and the add and edit series
bodies expect a non-default series type and monitor with distinct quality
and language profile ids so the two lookups cannot be transposed.
Radarr handler tests never asserted which block the search and filter keys
push, so transposing searching_block and filtering_block in a
TableHandlingConfig went undetected. Add a search key and a filter key test
for each of the three handlers that configure them: library, collections and
history.
No test moved a table or list selection off index 0 either, which made
current_selection() and items[0] indistinguishable. Select a non-zero index
in the tests covering every handler that reads a selected row, enriching
fixtures with a second, distinct entry where only one existed.
Several fixtures also asserted against a default value, so the expected
result was indistinguishable from an unset field. The system details tasks
now use two distinct non-default task names, and the add movie body expects
a non-default monitor and minimum availability.
A collection movie's overview renders in a popup that does not scroll,
so a long overview overflows with no way to read the rest. The popup now
scrolls.
The popup owns its own ScrollableText, built from the selected collection
movie's overview when it opens and torn down on Esc so a reopened popup
starts at the top. Scrolling mutates only that owned state. The
background renders the Collection record's overview and the movies
table, neither of which carries a scroll offset, so they cannot scroll
along with the popup. The popup also gains the loading and absent
fallbacks the rest of the tree already has.
Give the collection movie fixture an overview distinct from its
collection's, and a multi-paragraph one, so the snapshots can tell the
two records apart and a scrolled popup is distinguishable from an
unscrolled one.
A series' overview renders as one line of a description box that does
not scroll, so a long overview overflows with no way to read the rest.
Pressing v on the series details Seasons tab now opens a popup holding
the full overview.
The popup owns its own ScrollableText, built fresh from the series'
overview when it opens and torn down on Esc so a reopened popup starts
at the top. Scrolling mutates only that owned state. The background
paragraph renders the Series record's overview, which carries no scroll
offset, so it cannot scroll along with the popup.
Cite the existing view keybinding from the series details context clues
and give the popup its own clue set so it does not inherit the hints of
the view underneath.
Enrich the shared series fixture to a multi-paragraph overview, and the
raw response fixture that encodes the same series, so a scrolled popup
is distinguishable from an unscrolled one.
An artist's overview renders as one line of a description box that does
not scroll, so a long overview overflows with no way to read the rest.
Pressing v on the artist details Albums tab now opens a popup holding
the full overview.
The popup owns its own ScrollableText, built fresh from the artist's
overview when it opens and torn down on Esc so a reopened popup starts
at the top. Scrolling mutates only that owned state. The background
paragraph renders the Artist record's overview, which carries no scroll
offset, so it cannot scroll along with the popup.
Cite the existing view keybinding from the artist details context clues
and give the popup its own clue set so it does not inherit the hints of
the view underneath.
Enrich the shared artist fixture to a multi-paragraph overview, and the
blocklist response fixtures that embed the same artist, so a scrolled
popup is distinguishable from an unscrolled one.
Dark-Alex-17