Skip to content

fix: resolve the v1.3 RC QA blockers, high and medium findings - #402

Merged
smith-and-web merged 97 commits into
mainfrom
fix/v1.3-rc-qa
Sep 24, 2026
Merged

smith-and-web merged 97 commits into
mainfrom
fix/v1.3-rc-qa

Conversation

@smith-and-web

Copy link
Copy Markdown
Owner

Summary

Fixes every Blocker, High and Medium finding from the v1.3 RC QA report (prds/v1.3-rc-qa-report.md, local only), bumps the version to 1.3.0 and adds curated 1.3.0 release notes. Each finding comes with a regression test that fails without its fix. The work went through four high-effort code reviews, one after each section and a final one over the whole branch, plus two short follow-up passes on the review fixes.

Blockers

  • B1: re-clicking the active Beats/Page button no longer rebuilds prose from the stale copy, on both the frontend and in switch_scene_editor_mode.
  • B2: Page view is keyed per scene, so undo can't reach into the previous scene.
  • B3: version 1.3.0 in all five manifests and lockfiles.
  • B4: SECURITY.md now describes the updater and feedback requests. The website fix lives in the website repo; see notes.

High

  • H1: novelWriter sync has per-field baselines, conflicts, "keep kindling" settlement and a pre-apply snapshot.
  • H2: Beat view restores failed-save drafts.
  • H3: emoji edits no longer store lone surrogates, and poisoned journals are repaired.
  • H4: a focus trap and Enter handling in the export and Help dialogs.
  • H5: imported reference HTML goes through safeProse.
  • H6: Mac Scrivener RTF keeps paragraphs, code pages, tabs, and drops images and placeholders.
  • H7: Scrivener Create New refuses an existing .scriv and publishes without overwriting.
  • H8: XML-invalid characters are stripped from DOCX, EPUB, .scrivx and .nwx.
  • H9: Longform :: lines are no longer deleted, and titled Plottr cards import.
  • H10: [Bug]: AppImage shows white screen on CachyOS (Arch Linux) - missing dependencies? #252 reopened, with attempted-fix wording in the notes.
  • H11: release.yml reads the committed changelog section and fails if it's missing.

Medium

  • M1: tiptap 3.31.3, devalue, rustls, crossbeam-epoch and quick-xml 0.41 for the advisories.
  • M2: the MCP plugin sits behind a default mcp feature that release builds disable, and its npm package is a dev dependency.
  • M3: a new release-build.yml rehearses the release bundle matrix.
  • M5: the library is backed up before a new version migrates it.
  • M6: backend lock enforcement.
  • M7: automatic revision drafts are capped, and snapshot restore keeps draft history.
  • M8: prose is flushed before export and sync.
  • M9: delete-existing only removes files kindling wrote.
  • M10: atomic classic exports.
  • M11: novelWriter import and sync fidelity.
  • M12–M28: export workspace reveal, SMF workspace DOCX, classic scene-type filtering, smart quotes and fonts, feedback errors, timeout and disclosure, surfaced errors, kept typed text, Longform vault containment, inert HTML parsing, keyboard reordering, dialog focus everywhere, menu shortcuts under modals, dark-mode paper, the editor's accessible name, locked segment state, and closable Draft history.
  • M4: curated release notes in CHANGELOG.md, with Known Issues taken from docs/release-triage-v1.3.md.

Related Issues

Relates to #252 (kept open until the reporter confirms the attempted fix on 1.3.0), #334, #335, #341.

Type of Change

  • Bug fix (non-breaking change that fixes an issue)
  • New feature (non-breaking change that adds functionality)
  • Breaking change (fix or feature that would cause existing functionality to change)
  • Documentation update
  • Refactoring (no functional changes)
  • CI/Build changes
  • Other (please describe):

Checklist

  • I have read the Contributing Guidelines
  • My code follows the project's code style
  • I have added tests that prove my fix/feature works
  • All new and existing tests pass locally
  • I have updated documentation as needed
  • I have added relevant labels to this PR

Local gates on the final commit:

  • Rust: cargo test --all-features 618 passed. cargo clippy -D warnings is clean with --all-features and --no-default-features, and fmt is clean.
  • Frontend: vitest 1,016 passed. Coverage is 98.64 statements, 94.91 branches, 98.87 functions and 99.18 lines.
  • svelte-check and eslint show 0 errors. Prettier and check:design-system pass, and npm audit --omit=dev reports 0.
  • IPC: 166 of 166 commands registered.
  • commitlint passes for every commit.

Screenshots/Recordings

None. The UI changes are covered by component tests; see the manual checks below.

Testing Instructions

I haven't driven the running app. Before tagging, please do a manual smoke test on a built app (QA report §7.9):

  1. Upgrade from an installed 1.2.0 with a real library. Check that backups/kindling-*.db appears in the data folder and that projects open.
  2. Re-click the active Beats/Page tab (B1), and press Cmd+Z after switching Page-view scenes (B2). Prose must be unchanged.
  3. Export workspace → Show in folder. Send Feedback with Wi-Fi off: it should time out with a readable error.
  4. Double-click a .kindling-review file. Quit from the Dock with a fresh edit.
  5. Check that the new Release Build workflow passes on this PR. It is the first real run of tauri build … -- --no-default-features.

Additional Notes

  • Schema: two additive, idempotent tables, novelwriter_sync_baselines and novelwriter_split_documents, both covered by migration tests. No other schema changes.
  • Behaviour changes to note:
    • Existing Plottr projects will be offered their previously dropped titled cards as new scenes on the next sync.
    • Longform key:: value lines are read as fields only when the key is lowercase, the value is links only, or the key is one kindling maps. Other lines stay in the prose.
    • The first-visit guidance tip now blocks menu shortcuts while it's showing.
  • Website: the privacy page still says the app makes zero network requests. The fix is already on the website repo's unmerged press/editorial-restructure branch and should ship with this release.
  • Known and left as they are:

🤖 Generated with Claude Code

smith-and-web and others added 30 commits September 23, 2026 21:54
…after a scene switch

Clicking the already-selected Beats or Page button rebuilt the scene's
prose from the other view's stale copy. The switch is now a no-op when
the scene is already in the requested mode, in both the frontend and
switch_scene_editor_mode.

Page View kept one TipTap editor across scene changes, so undo could
step back into the previous scene's text and autosave it into the
current scene. The page editor is now keyed on the scene it was loaded
with, giving each scene a fresh undo history.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The changelog job regenerated notes at tag time, when package.json already
matches the tag, so it emitted an empty section and the release fell back
to "See CHANGELOG.md". It now reads the tag's section from the CHANGELOG.md
committed at the tag and fails if there isn't one.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… on import

Importing a Scrivener project mangled the prose of every scene:

- Mac Scrivener writes each paragraph break as a backslash followed by a
  newline. That was ignored, so every scene imported as one paragraph.
- Curly quotes, apostrophes, dashes and ellipses stored as cp1252 bytes
  (\'92, \'93, \'97, ...) became invisible control characters. They now
  decode as cp1252, honouring \ansicpg when it names Latin-1.
- Tabs were dropped, running words together.
- Images embedded by Windows RTF writers leaked pages of hex into the
  prose. Picture, object, header/footer, list-template and other
  non-text destination groups are now skipped.

The same pass fixes neighbouring RTF handling: Unicode characters with an
ANSI fallback were imported twice and \ucN skip counts were ignored,
characters outside the BMP were lost, bold/italic set inside a {group}
leaked past its end, an escaped brace inside a skipped group could hide
the rest of the document, \line and Cocoa's line separator now become
soft line breaks instead of new paragraphs, and a content.rtf containing
raw 8-bit bytes (not UTF-8) now imports instead of arriving empty.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Importing a Longform project treated any scene line containing "::" as a
Dataview inline field and removed it from the prose, so sentences such as
"The odds were 3::1" or "Night falls:: the wait begins" silently vanished.

A line is now only a field when it is exactly "key:: value" with a single
field-name key (a letter followed by letters, digits, "_" or "-"). Keys
Kindling maps (pov, characters, setting, ...) are still recognised without
a space after "::"; other keys need one, so text like "std::vector" stays.
Ambiguous lines, including multi-word keys, list items and quotes, are kept
as prose.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Importing a Plottr file dropped every card that had a title but an empty
description, with no warning. Plottr's Summary-line cards are usually like
this, so the Hamlet sample lost 5 of its 25 cards.

A card now imports as a scene whenever it has a title or description text;
a titled card with no description becomes a scene with no synopsis or
beats. Only completely blank cards are still skipped. Cards in the same
beat and row are ordered by storyline, so a Summary card leads its act
instead of depending on the order cards happen to appear in the file.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Prose pasted from Word carries U+000B for a Shift+Enter line break, and
imports can bring in other C0 control characters. XML 1.0 forbids them,
and escaping `&<>"'` does not remove them, so a manuscript containing one
exported to a DOCX that Word reported as "unreadable content" and an
EPUB that readers refused to open. Classic DOCX, treatment DOCX, classic
EPUB and the export workspace's DOCX and EPUB were all affected.

Every DOCX now packs through one sanitiser that cleans each XML part, and
every EPUB text node, title and metadata field goes through the same
sanitiser. A manual line break (U+000B) or form feed (U+000C) in prose
becomes a real soft line break, as `<br>` does; in titles and metadata it
becomes a space. Other forbidden characters, including U+FFFE/U+FFFF,
carry no visible meaning and are dropped.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Exporting to Scrivener with "Create new" wrote straight into whatever
.scriv bundle was at the chosen path. The save name defaults to the
project's name, which is also the name of the .scriv it was imported
from, so saving next to the original and accepting the "Replace?"
prompt replaced the writer's binder, including Research and Notes that
kindling never imports, with no backup.

Create New now refuses when anything already exists at the destination,
and the dialog explains why: choose a new name, or use Update Existing to
write prose into that project. The check runs before the pre-export
snapshot. The bundle is also built in a hidden folder beside the
destination and moved into place only when complete, so a failed export
no longer leaves a half-written .scriv behind, and a project that appears
at the destination mid-export is still never replaced.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
When a beat's save failed (database busy, read-only volume), collapsing and
reopening the beat, or leaving the scene and coming back, remounted the
editor from the last successfully saved prose. The banner still said "your
text is still here", but the editor showed the older text, and the next
keystroke saved that older text over the recoverable draft.

Beat View now mounts, previews and counts words from the newest unsaved
text for the beat (the open draft, then a failed or in-flight save), the
same way Page View restores from the recovery queue.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Replacing one emoji with another that shares its first UTF-16 half (for
example 😀 with 🙂, or changing a skin tone) stored a suggestion containing
only half of each character. The backend rejects that text, so every save,
feedback export and recovery export of the review failed until the edit was
undone, and the local recovery journal replayed the failure on reopen.

Tracked changes are now widened to whole characters before their text is
stored. A recovery journal already holding half a character is repaired on
load: its suggestions are re-derived from the reviewed manuscript, keeping
their ids and discussion.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Reference descriptions from imported files (yWriter bios and goals,
Longform descriptions and others) were inserted into the References panel
and the reference type dialog as raw HTML. A crafted project file could
include a page refresh, a sign-in form or an invisible full-window link,
which the app's content security policy does not block, and expanding the
reference would expose it.

Descriptions now pass through the same allowlist the export preview uses,
which keeps paragraphs, line breaks, bold, italics, lists and headings and
drops everything else. The allowlist moves to a shared safeHtml module.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Opening Export (Cmd+E) while writing left focus in the manuscript, so
typing and Enter kept editing the prose behind the dialog. Enter anywhere
in the export dialog also started an export, even on Cancel or Choose...,
and Escape reached the scene panel behind it. About, Quick Start and Send
feedback behaved the same when opened from the menu.

These dialogs, and the Scrivener match preview opened over the export
dialog, now move focus inside when they open, keep Tab and Shift+Tab
inside, keep key presses from reaching the page behind, close only the
topmost dialog on Escape, and return focus to where it was when they
close. In the export dialog Enter exports only when focus is not on a
button, link, select or text area; those controls handle Enter themselves.

The behaviour lives in a reusable modalFocus action
(src/lib/utils/modalFocus.ts) so the remaining div dialogs can adopt it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ter project files

Titles carrying control characters produced a .scrivx or nwProject.nwx that
Scrivener and novelWriter refuse to open. Both writers now use the same
sanitiser as the DOCX and EPUB exporters.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
novelWriter sync had no record of what was last synced, so every edit made
in kindling since import was previewed as an incoming change carrying the
old novelWriter text. Choosing All and applying silently reverted those
edits, with no snapshot to recover from.

kindling now records the novelWriter text of each synced field at import
and after every applied sync (new additive novelwriter_sync_baselines
table). Sync only offers a change when novelWriter actually changed since
that baseline; kindling-only edits are never offered as incoming. When both
sides changed, or a project was imported before baselines existed, the
change is shown as a conflict: All skips it and it must be chosen
individually, and reimport never applies it. Declined changes keep their
old baseline so they are offered again next time.

Applying a sync that replaces prose now takes an automatic "Before sync"
snapshot first.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…fari 16.4

The check that repairs emoji halves in a recovered review used a regular
expression lookbehind. The system WebKit on macOS releases before Safari
16.4 cannot parse one, so the whole editorial module failed to load there
and every screen that imports it broke. The check is now a plain character
scan with the same results.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… text

A collapsed beat previews its newest unsaved text, but the preview only
refreshed when something else on the page changed. A save that failed,
finished or was discarded elsewhere could leave the preview and word count
showing the wrong version. The save queue now notifies views that ask to
observe it, and Beat View builds its lookup of unsaved beat text once per
change instead of searching the queue for every beat on every render.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ng text

A conflict the writer left unticked came back on every sync, and on
projects imported before sync baselines every locally edited field stayed
a conflict until the writer accepted novelWriter's text over their own.

Applying a sync now settles each unticked conflict as "keep kindling": the
kindling text stays and the conflict is not offered again unless
novelWriter changes that field again. Unticked non-conflicting changes are
still offered next time. The dialog says plainly that leaving a conflict
unticked keeps the kindling version, and Apply ("Keep kindling versions")
is available even when nothing is ticked. Reimport, which shows nothing,
never settles conflicts.

Apply now compares the project once (again only when it added items) and
writes a sync baseline only when its value actually changes, instead of
rewriting every field's prose on every sync.

The automatic "Before sync" snapshot is removed again when the apply fails
or replaces no prose, so failed or retried syncs no longer leave snapshots
behind.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…switches

Deleting an unlocked chapter silently deleted any locked scene inside it,
along with its beats and prose. Part deletion already refused this; single
chapter deletion now applies the same rule, archived locked scenes included.

The backend now also refuses the other changes the app already blocks for
locked scenes and chapters, so a stale or scripted call can't bypass the lock:

- reordering: a locked scene or chapter keeps its slot; others can still be
  rearranged around it
- moving a locked scene to another chapter, or into a locked chapter
- renaming a beat of a locked scene
- switching a locked scene between beat and page view (re-selecting the
  current view still succeeds)
- changing the planning status of a locked scene or chapter
- archiving a locked scene, or a chapter that is locked or holds one
- converting a locked chapter to or from a Part

The sidebar now shows these refusals (and other delete, reorder, archive,
planning and Part-conversion failures) instead of appearing to do nothing.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
The export workspace's "Open export" button always failed after a
successful export: it called openPath, which the app's opener permission
does not grant, so the request was refused every time. It now reveals the
file in Finder/Explorer, as the classic export's success dialog already
does, and is labelled "Show in folder" to match. If the folder can't be
shown, the reason appears in the workspace's error notice rather than only
in the console.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…er in dark mode

In the dark theme, an HTML export with built-in styling turned off
previewed as black text straight on the dark desk (about 1.15:1), which
was unreadable. The preview now falls back to manuscript paper and ink
whenever the export sets no sheet colours of its own; styled exports keep
their own look.

The draft history comparison showed both versions' prose on the dark
chrome rather than on paper, and added words were green text measuring
4.33:1. Both versions now sit on the manuscript sheet, and changed words
keep manuscript ink; the tinted wash and the strike-through or underline
mark what changed.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Screen readers announced the main prose editor as an unnamed editable
area, so a writer moving between the toolbar, beats and the page couldn't
tell where they were. The editor is now a named multi-line text box,
"Scene prose" in Page view and "Beat prose" inside a beat, and it reports
when a locked scene makes it read-only.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Locking a scene disables the Beats/Page switch, and the app-wide
disabled style then stripped the selected fill from the active segment,
leaving both options the same colour as their track. Nothing showed which
view the scene was in. The chosen segment now keeps a raised fill and a
ring in the disabled ink, so it still reads as selected in both themes
while looking unavailable.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… reload fails

In the review workspace, Draft history could not be dismissed while the
workspace was still reloading (for example straight after restoring a
draft): the close button and Escape were silently ignored. If reloading
the manuscript on close failed, the dialog stayed open and every further
attempt failed the same way, trapping the writer in it.

Draft history now always closes at once. The saved drafts are reloaded
after any work already under way finishes, so the two can't interleave,
and a failed reload is reported in the workspace's error notice.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A failed submission showed "[object Object]" because the backend error
crossed IPC as a tagged enum; it now arrives as a readable message. A
black-holed network left the spinner running forever; requests now time
out after 15 seconds. The dialog now says a submission includes the kindling
version, operating system and language, and nothing from the manuscript.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Every project lives in one kindling.db, and a new version migrates it in
place at startup with no copy to fall back on. The first launch of each new
version now writes a consistent copy to backups/ in the data folder (the
newest three are kept), and a failed migration names that copy.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Release bundles (universal macOS, ubuntu-22.04 AppImage/deb/rpm, Windows
MSI/NSIS) were only built by release.yml at tag time. A new workflow builds
the same matrix unsigned on pull requests that touch packaging, weekly and
on demand.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…page

Two gaps remained in Scrivener RTF import:

- A picture pasted into a Mac Scrivener document left a stray "¬" in the
  prose. Cocoa writes the image as a NeXTGraphic group followed by a
  U+00AC placeholder; the placeholder right after an attachment is now
  dropped, as is U+FFFC (object replacement) anywhere.
- Documents in a code page other than Western European imported with
  the wrong letters (Cyrillic, Greek, Central European, Japanese, Mac
  Roman, ...). Text is now decoded with the code page the file declares
  (\ansicpg, or \mac), falling back to Windows-1252. Raw 8-bit bytes in a
  non-UTF-8 content.rtf and \'hh escapes use the same code page, and
  consecutive escapes are decoded together so double-byte characters
  survive.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
When a scene's synopsis was already set by Kindling's own metadata, a
Dataview line such as "synopsis::New text" (no space after "::") was
treated as prose and appeared in the scene text. A key Kindling
recognises now always marks the line as metadata, even when its value is
ignored because the field is locked.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
smith-and-web and others added 18 commits September 24, 2026 07:39
…leted

Sync read a multi-heading document whole, the way kindling 1.2 imported
it, whenever the project had the scene from its first heading but none
from its later headings. A document imported one scene per heading, whose
later scenes the writer then deleted, matched that too, so the first scene
was offered the whole document's text and headings.

kindling now records each document it reads one scene per heading, at
import and after every sync (a new additive table), and only reads a
document whole when it has no such record, which only a 1.2 import can
lack. Deleted later scenes are simply offered back as new scenes.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Importing a Longform scene silently deleted any line shaped like
"Word:: text" whose key Kindling does not use, so sentences such as
"Meanwhile:: the ship sank." vanished, along with genuine custom
Dataview fields like "mood:: tense".

Only keys Kindling maps (pov, characters, setting, status, synopsis and
the other reference keys) are now treated as metadata. Scenes have
nowhere to store other fields, so those lines stay in the prose where
the writer can see them.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Moving a scene to another chapter wrote only the scene's own position. The
target chapter's scenes weren't shifted and the source chapter kept a gap.
Two scenes could then share a position and come back in an arbitrary order,
which made later reorders of that chapter fail with a locked-scene error.

The move now renumbers the scenes in both chapters in one transaction, so
positions never collide. As when reordering, a locked scene keeps its place:
placing a scene at or before a locked scene in the target chapter is refused
with a clear message, rather than silently put somewhere else. Leaving a
chapter closes the gap, as deleting or archiving a scene does. Positions past
either end of the chapter are clamped to that end.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…re's no author

With no author name set, the classic Word exporter's running header read
" / MY NOVEL / 2", with a stray leading slash. The export workspace
already dropped the author part. Both now compose the header text in one
shared place, so it reads "MY NOVEL / 2" without an author and
"Surname / MY NOVEL / 2" with one.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Manuscript typography removes the spaces around an em dash, but across
inline markup it only looked for a literal em dash at the start of the
next piece of text. "word <em>--then</em>" therefore exported as
"word —then", while "word --then" gave "word—then". Two spaces before a
typed dash also left one behind.

The next piece of text is now judged in its normalised form, where --
and --- are em dashes and spaces before a dash are dropped. Double
spaces are collapsed before the dash rule runs. Every way of typing the
dash, with or without italics, now gives "word—then".

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…rose

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…is pruned

When saving pruned the draft on the comparison side of Draft history, it
fell back to the nearest older draft, which could be the very draft on the
other side, leaving an empty comparison. The comparison side now moves to
the nearest newer draft, neither side takes the other's draft, and the
current prose is the last resort. The notice names both drafts now shown,
and says nothing about a comparison when no drafts are left.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
… is added

kindling 1.3 records which novelWriter documents it reads one chapter or
scene per heading, so documents imported whole by 1.2 can keep that shape.
It only recorded documents that already had several headings, so a
one-heading document imported by 1.3 that later gained a heading in
novelWriter looked like a 1.2 import: the new section was offered as more
text for the first scene instead of as a new scene.

Every Novel document read per heading is now recorded at import and after
each sync, whatever its heading count. Documents imported whole by 1.2 are
still unrecorded and keep their one-scene shape.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Applying a sync checked the reviewed changes by parsing the whole
novelWriter project and comparing it with kindling while holding the
database lock, then released the lock, took the snapshot, and parsed and
compared everything again to apply.

Apply now parses and compares once. After checking that the reviewed
changes are current it takes the pre-sync snapshot on the connection it
already holds, only when those changes replace prose, then writes. A stale
or refused apply still writes no snapshot, and one whose writes fail
removes it again.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…alues

The previous fix kept every unmapped field line in the prose, which
brought Dataview breadcrumbs such as "up:: [[Chapter One]]" into scenes
and turned each link into a character. It also still ate sentences
whose first word happens to be a key Kindling maps, such as
"Crew:: all hands stood silent.", and filed them as organizations.

A line is now a Dataview field only when its key is lowercase-initial
(the Dataview convention) or its value is nothing but wikilinks. Mapped
fields are applied as before; other fields are consumed as metadata,
as in 1.2, without creating references. Anything else, including
"Meanwhile:: the ship sank." and "Crew:: all hands stood silent.",
stays in the scene prose.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
A scene move refused because it would shift a locked scene always said
"Cannot place the scene before the locked scene … Choose a later position".
That was wrong when the scene was being moved from before the locked scene
to after it in the same chapter. The message now names the locked scene
and says the move would shift it, which is true in both directions.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ve helper

reorder_scenes now calls reorder_scenes_in_transaction instead of repeating
its UPDATE loop. db::move_scene_to_chapter had become test-only, because the
move command plans and applies the move itself so it can check locks in
between. It is removed, and its two tests now call plan_scene_move and
apply_scene_move directly.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…data

The key-case rule made a capitalised mapped key with a plain value, such as
POV:: Zoe or Status:: Draft, import as prose, where 1.2 read it as
metadata. A capitalised key is now a field when kindling maps it and the
value does not read as a sentence, so Crew:: all hands stood silent. still
stays in the prose.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…commas in link aliases

A capitalised mapped key such as Synopsis:: She arrives at the dock. was
kept as prose when its value ended like a sentence, so the synopsis was
lost. Capitalised keys kindling maps are metadata again, whatever the value.
A breadcrumb whose link alias contains a comma, or an embed, is now
recognised as links-only instead of creating a character reference.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Comment thread src/lib/utils/outlineMove.test.ts Fixed
smith-and-web and others added 6 commits September 24, 2026 10:35
The helper removed only the first asterisk from a row spec. Replace them
all (CodeQL: incomplete string escaping). Suggested by GitHub Copilot
Autofix.
ProseMirror 1.42 marks every Escape inside the editor as handled, and the
scene panel had started ignoring Escapes another control already handled, so
Escape no longer collapsed an expanded beat. An Escape from the editor now
counts again; one handled by a dialog or another control still doesn't.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
plist 1.8 (through tauri) pulled in quick-xml 0.38 (RUSTSEC-2026-0194,
RUSTSEC-2026-0195). plist 1.10 no longer depends on it.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
…ugin's build tree

The last copy of quick-xml before 0.41 is a Linux build-dependency of xcb,
reached only through the debug-only tauri-plugin-mcp, which release builds
leave out. It parses xcb's bundled protocol files at build time, never user
input.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Lockfile-only updates within the existing ranges for vitest, eslint, the MCP
test server and WebdriverIO tooling (undici, hono, qs, postcss, js-yaml,
brace-expansion and others). No production dependency changes.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Development-only advisories are still reported but no longer fail the job:
extract-zip, reached through WebdriverIO's browser downloader, has no fixed
release and never runs in the app.

Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
@smith-and-web
smith-and-web merged commit ceaff78 into main Sep 24, 2026
16 checks passed
smith-and-web added a commit that referenced this pull request Oct 30, 2026
…for macOS 27.2 (#421)

* test(qa): wait for the editorial workspace and export pane before acting

Suite 17 failed on main since ceaff78 (#402): closing Draft history now
reloads the local scenes under the workspace's busy flag, and action()
silently drops any manuscript-menu action started while it is set. The
suite reopened the menu and clicked "Review packages and rounds…" before
that reload settled, so #package-title never appeared. It now waits for
the Editor mode select to re-enable (it is disabled={busy}) first. The
same dropped click is reachable by a fast user because the menu items are
not disabled while busy; that is an app follow-up, not papered over here.

Suite 19 raced the export source pane: 19-03-export-html asserted on
.source-preview before the pane had regenerated after the mode and tab
change, passing in light and failing in dark by timing alone. It now
waits for the pane's content before capturing.

The runner also pins the scene view's planning callout. It is gated only
by kindling:tooltipSeen:planningStatus, which setVisualPreferences does
not touch, so captures depended on whether the QA profile had dismissed
it. Each variant switch now removes the key (fresh-profile state); the
saveLocal/restoreLocal guard puts the user's value back afterwards.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* chore(qa): re-accept the visual baselines after the macOS 27.2 update

Every capture differed from its baseline by 0.6 to 3 percent of pixels.
The baselines were captured on macOS 26.6.2; macOS 27.2's WebKit renders
Inter and Fraunces about 2 CSS px taller per line, so every card, row
and dialog sits a few pixels lower. The previous acceptance commit
(72266be) renders pixel-identically to main on this OS, so no source
change is involved and no bisect was needed.

All 147 masters were inspected against Expect as contact sheets beside
the previous masters with pixel-diff overlays, then at full size for the
largest differences. Besides the metric shift, the accepted set picks up
changes already on main (the scene planning callout, the feedback
dialog's privacy sentence, the outline opening the first act with its
filter row, the 1.3.1 version label and the export preview's Standard
Manuscript Format running head from pull request 402). The References
aria-required-parent finding no longer reproduces (08-02 passes axe in
all variants), so its knownFindings entries are removed.

Verification against the refreshed set gave 147/147 matched images, 147
checkpoints passed, exit 0, in 248 seconds. The README now explains how
to tell OS drift from a source regression before bisecting.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

* ci(workflows): pin the Linux E2E job to ubuntu-24.04

GitHub moved ubuntu-latest to 26.04 on 2026-10-30. There apt has no
webkit2gtk-driver (it is replaced by webkitgtk-webdriver), so the
"Install system dependencies" step exits 100 before anything runs. The
job last passed on the 24.04 image on 2026-10-20 with no workflow
change in between. Pinning restores that image; moving to 26.04 and
the newer WebKitGTK line is a deliberate change for its own PR, since
tauri-driver has not been validated against it.

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

---------

Co-authored-by: Claude Fable 5.1 <noreply@anthropic.com>
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

2 participants