On Windows, a network path from note content that lies outside the vault
and outside any folder the user opened is refused before the filesystem is
asked about it, in the backend path checks and in the frontend resolvers.
Other platforms are unchanged.
Frontmatter is parsed through a YAML engine with limits on size, aliases
and nesting. Frontmatter outside those limits is read as a note without
metadata instead of being handed to the unbounded loader.
A note that cannot be converted for the rich editor no longer leaves the
previous note's content in place. The editor is emptied and locked, saving
is blocked, and source mode shows the real content. Programmatic document
changes no longer leave the next user edit untracked.
Paths from a WebDAV listing and from the stored sync state are now checked
before they are used as local file paths, and other spellings of the app's
internal folder are never synced.
A local file failure during WebDAV sync surfaced as a bare OS error such as
"Operation not permitted (os error 1)", with no file and no hint that it
was local. Local failures now name themselves, the operation and the file,
and requests that get no answer name the operation and remote path.
Refs #249
The title was the only part of the header that could shrink, down to a few
characters or nothing. It now keeps a minimum width and the toolbar buttons
move to a second row when there is no room. Unchanged when there is room,
and on mobile.
Refs #246
Ctrl+0 reset the editor font to a fixed 14 px, because zooming overwrote
the size chosen in Appearance. That choice is now kept separately and
Ctrl+0 returns to it. For existing installs the size in use at first
launch becomes the preference.
Refs #272
Text formatted with a trailing or leading space was saved as **bold **rest,
which Markdown does not read as bold, so it showed as literal asterisks
after reopening. Edge spaces now go outside the markers and a mark over
only spaces is not written. Formatting around an inline code span is
dropped when a letter or digit touches it, where it would not render.
Refs #240, #276
Backspace at the start of a collapsible section's body did nothing, so the
caret could never leave it and a section on the first line of a note could
not be removed from the keyboard. It now steps back into the summary, where
further Backspaces delete the text and unwrap the section. A section with
nothing in it is removed.
Refs #278
A round-trip audit of 1590 notes showed four things lost the first time a
note was edited. A code block without a language was saved as "text".
Link titles were dropped. Table column alignment was reset. Bold or
italic around inline code was removed.
All four now survive loading, editing and saving. Formatting that covers
only part of a code run is still dropped, since it cannot be written
inside a code span.
A line with thousands of unclosed [label]( starts took about 1.1 s at
56 KB, because each start read to the end of the line for its closing
parenthesis. The legacy link rule now skips starts beyond the last
closing parenthesis and bounds the failed searches per block.
The image toolbar stayed open after an undo and its buttons then changed
a different image. It now closes when its image is no longer at the
remembered position.
Images separated by a blank line load with an empty paragraph between
them, so the Grid button never found two neighboring image blocks and
did nothing in an existing note. Grouping now spans a single empty
paragraph between image blocks and takes the one above the first image,
which would otherwise be saved as a marker line above the grid.
Ungroup gave each image its own paragraph but no blank line, so the
images were saved directly under the preceding paragraph and became part
of it on reload. Top-level grids now ungroup into images a blank line
apart, the reverse of grouping them.
Refs #254
For content copied from another app, WebKitGTK's navigator.clipboard.read() often
returns nothing, or only text/plain, while Ctrl+V still receives the HTML. The
context menu then pasted plain text and lost bold, links and other formatting.
Whatever the web clipboard leaves out is now filled in from the native clipboard.
The new read_clipboard_contents command (replacing read_clipboard_text) returns
both text and HTML through arboard. The result is replayed through the same paste
handlers as Ctrl+V. That includes the WebKitGTK image fallback for an <img>-only
clipboard. The direct native image read only runs when there is no text or HTML.
Images now keep their CommonMark title as hover text through load, edit and
save: ``. Link titles are still dropped, as
before.
Parsing:
- helix_images no longer strips `title` from images (links unchanged).
- An image whose destination and title are valid CommonMark now uses
markdown-it's own image rule instead of the legacy-link rule. Previously
any space before a title sent the image down the legacy path, which can't
handle escaped quotes (`"Say \"hi\""`) or `(title)` titles and mangled
them into the URL.
- Genuinely legacy forms (spaces in the path, `%20"title"`) still go through
the legacy rule, which now also keeps the image's title.
Saving: `serializeImageTitle()` writes ` "title"` with CommonMark backslash
escapes (`\"`, `\\`, and `\&` only before entity-like text). The parser
undoes these exactly, so titles survive any number of saves unchanged.
UI: a "T" button on the image toolbar opens a small dialog to set or clear
the hover text.
Tests: tests/image-titles.test.mjs covers exact round trips for tricky titles
(parentheses, &, entities, quotes, backslashes, markup), no growth over
repeated saves, the written form, titles from other apps, legacy
destinations, and links still dropping titles. The upstream assertion that
images render without `title=` is updated to the new contract.
Implements #254. Several images can now sit side by side in a grid or masonry
layout with 2 to 4 columns. The syntax is plain Markdown images inside a
`:::` container, so other renderers show the images stacked:
:::img-grid cols=3 masonry


:::
Parsing is a markdown-it block rule (`helix_img_grid` in markdownParser.ts),
not a text pre-pass:
- Grid syntax inside code blocks is never converted, because code blocks
claim those lines first.
- A block converts only if every line is exactly one image (CommonMark image
syntax, so titles, parentheses and <angle> destinations parse as for any
image). Unknown options, out-of-range columns, captions, linked images,
blank lines or a missing closer leave the whole block as ordinary Markdown.
- Each grid image is an ordinary inline image token, so it gets the same
source resolution, `|size=` handling and title policy as other images.
Saving writes the grid as a normal block through joinMarkdownBlocks, with a
blank line before it. The images serialize with the same code as images in
paragraphs, one per line. The header is written in canonical form.
Editor (src/lib/editor/extensions/imageGrid.ts):
- `imageGrid` node (`image+`) whose renderHTML includes data-img-grid, so
copied grids paste back as grids
- header controls: cycle columns, grid/masonry toggle, ungroup
- image toolbar: "Grid" groups the image with neighboring image-only blocks;
arrow buttons move an image left or right inside its grid
- drag-and-drop reorder within and between grids; an emptied grid is removed
- `/image-grid` slash command: multi-file picker that inserts a grid, using
the origin-guarded asset saving
- mobile: 3-4 column grids collapse to 2. Print: compact 2-column rows,
controls hidden
Tests (tests/image-grid.test.mjs) cover byte-identical round trips, sizes,
code-fenced examples, no partial conversion, titles and angle-bracket
destinations, repeated saves, trailing closers, nesting, reloading next to
other blocks, renderHTML/parseHTML agreement, grouping and reorder targets.
On Linux, the context-menu Paste used its own navigator.clipboard.readText()
flow. On WebKitGTK it lost text copied from other applications. When a web
image copy also carried text, it pasted the URL instead of the bitmap. It
also skipped ProseMirror's paste handling, so things like callout paste and
the pasted-HTML style cleanup never ran.
The context menu now reads the clipboard with navigator.clipboard.read(),
classifies each format, and replays the result as a paste event on the
editor DOM. That runs the same path as Ctrl/Cmd+V: ProseMirror's paste
handler, every handlePaste prop (file, callout, link), transformPastedHTML,
the clipboard parsers and TipTap's paste rules.
Only real attachments (image/*, application/pdf) become files. Other text
formats such as text/uri-list stay text, and browser/OS metadata formats are
ignored, so a copied link no longer becomes a pasted-0.uri-list attachment.
A link offered only as a URI list pastes as its first URI.
When the web clipboard is unavailable (WebKitGTK bug #218519), Paste falls
back to native arboard reads: the image first, then text through the new
read_clipboard_text command. The text is replayed through the same paste
path. Every await re-checks the editor origin, as upstream's paste code does.
The classification lives in src/lib/utils/clipboard-paste.ts and has
regression tests.
Closes#259
Switching quickly between notes in the rich editor could leave a frozen
caret from the previous note next to the real one. The outgoing native
selection is now cleared before the note content is replaced, including
for very large notes. Reproduced on Linux and gone with the change.
Refs #281
Strip a legacy link title with a single scan instead of a regular
expression that is quadratic on a long run of spaces. Widen the done
checkbox's tap area on phones without changing its look, and say "1
unchanged line" in Version History when only one line is folded.
For notes over 100,000 characters, show the first screen at once and build
the rest of the document in growing slices. Until it is complete the
document cannot be edited: a user action, saving, or switching modes
finishes the build first, and switching notes cancels it. Smaller notes load
as before.
The Overdue filter now lists only tasks that are late and still open, even
with Hide done off, and a completed task's due date no longer uses the
overdue colour. Today and the calendar are unchanged.
Keep a second paragraph, code block or quote inside its list item when a
note is saved, indented under the item. Import attachment links that follow
a stray backtick and images indented by four spaces. Keep a folded callout
that was revealed from the Outline open while it is edited.
Stop treating the editor's first normalization of a freshly loaded note as
an edit, so opening a note no longer saves or reformats it. Join serialized
blocks in linear time, so autosaving a large note no longer freezes the page.
Write a literal $ as typed unless plain dollars could create, remove or
re-pair an inline formula. Save portable image paths when a centered,
aligned or styled block is written. Percent-encode a newly stored
attachment's path before it is inserted, and keep a Windows drive letter
literal in paths outside the vault. Keep whitespace-only code spans in
simple list items, task items and table cells.
Compare Version History in bounded time with a synchronous line diff and
an edit limit. Beyond the limit, shared lines at the start and end stay
unchanged and the rest is shown as removed and added.
Render the first soft break of a callout's opening paragraph as a newline
so the title and body split correctly again. Only turn [ ] and [x] markers
into task items in bullet lists, not numbered ones. Join serialized blocks
in markdown.ts so an empty paragraph before an image is written as a blank
line, and lists no longer gain an extra blank line, without collapsing blank
lines inside code blocks.
In Version History, compare the current note and snapshots after ignoring
leading and trailing blank lines, so unchanged last lines and a blank line
after the frontmatter no longer show as changes.
A vault kept in a dot-named folder such as ~/.notes was treated as hidden
by seven vault walkers, so All Notes, note titles for links, Graph View,
Tasks and rename link updates found no notes. Exempt the vault root from
the hidden-folder check, as the tag scan and search already do. Hidden
subfolders, .helixnotes and the existing hidden-name rules are unchanged.
Note lists, mobile listings and tag counts read a fixed 1-2 KB prefix, so notes
with longer frontmatter lost their title and tags. Read the complete frontmatter
instead, stopping after the closing delimiter (256 KB cap if unterminated).
Scan note titles in parallel from the frontmatter only, and build graph nodes and
links in one parallel read per note without keeping note bodies in memory. The
Ctrl+K link dialog reuses the loaded wiki-link title cache.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Protect file mutations, backup restoration, and WebDAV synchronization from data loss. Preserve editor drafts across concurrent saves, external changes, and window or vault transitions. Integrate toolbar customization inline and extend regression coverage.
- Show search results in the right notebook or All Notes list
- Scroll to the selected note and highlight matching text
- Keep unsaved edits safe and ignore old or cancelled searches
- Show save errors with a Retry button
- Run graph and vault statistics scans in the background
- Fix folder checks when counting vault files
- Keep desktop window-close code out of mobile builds
Replace the source editor with CodeMirror and reduce typing-time work.
Virtualize note lists, apply targeted refreshes, and batch search indexing.
Improve vault and in-note search, and bound graph layout work.
Make saves revision-safe and flush pending changes before navigation or close.
Preserve originating-vault ownership across async operations and detached windows.
Harden sync, attachment writes, and notebook/tag appearance updates.
Keep Trash previews read-only and skip restoring trashed notes.
Update AI model presets and isolate streamed responses by request.
Fix local image rendering and preserve portable image paths.
Include regression tests for the changed behavior.