The API rejects a save that changes the title while most of the note
content disappears with an English message, and clients showed it
verbatim in every locale.
- shared: add editor.saveState.errorReason.suspiciousOverwrite in all
four catalogs and getMemoSaveErrorReasonKey(), matching the error code
or the server message; the API reuses the exported code and message.
- desktop sync details: translate the item by lastErrorCode.
- Android: map the stored message to its zh-CN source text so the
localized components translate it.
- web editor: show the reason in the save status tooltip and label
instead of the bare "Could not save".
The automatic sync runner reset runningRef inside a finally block, after
awaiting refreshSyncQueueItems(). If reading the queue threw there,
runningRef stayed true and every later call returned early as if a pass
were still running, so automatic sync stopped until the app restarted.
A failing getMobileSyncRetryDelay() also surfaced as an unhandled
rejection from the void-ed run.
Move the settle step into settleMobileSyncPass(), where refreshing the
visible queue and reading the retry delay each handle their own errors
and the runner is always marked idle. Ordering is unchanged: refresh,
mark idle, run a pass requested meanwhile, otherwise schedule the retry.
This also removes the return inside finally reported by
no-unsafe-finally.
Co-authored-by: tianma-if <yingwaizhiying@gmail.com>
* refactor: resolve oxlint correctness findings without behavior changes
- Snapshot iterations that mutate their collection with Array.from()
instead of a spread (live childNodes, self-removing disposers, a Set
extended inside its own loop).
- Drop empty spread fallbacks, redundant regex escapes, Boolean() casts
inside negations, a rethrow-only try/catch and statement ternaries.
- Use startsWith/endsWith for anchored literal checks in the extension.
- List U+FE0F explicitly in the mobile attachment emoji class; the
matched set is unchanged.
- Fix the WeChat copy fixture, where \oint in a template literal lost its
backslash and the block formula never exercised \oint.
- Document the intentional control-character regexes (filename
sanitizing and input validation), the throwing provider stream test,
Astro's generated type reference and the mobile auto-sync finally
block with narrow inline exceptions.
* ci: run oxlint correctness rules on pull requests
Pin oxlint 1.88.0 and enable only its correctness category across apps/
and packages/, excluding build output, vendored code and the iOS app.
Findings are errors and the script passes --deny-warnings, so CI fails
on any finding. no-unused-vars stays off until unused code is removed
in a dedicated pass.
The shape-library trigger opened its menu on every pointerenter. When the
editor replaced the note list under a pointer that did not move, or when
the pointer only passed over the toolbar, the menu opened by itself and
stayed open. For mind maps and flowcharts that menu was modal, so the
back-to-list button could not be pressed until the menu was closed. On
touch screens pointerenter fires right before the pointerdown that
toggles the menu, so a tap could open it and close it again.
- Open the menu on hover only for mouse pointers.
- Make the mind-map and flowchart insert menu non-modal, as the
architecture library already is.
New mind maps, flowcharts and architecture diagrams always started with
Chinese labels, whatever language the interface used.
- createDefaultDiagramDocument takes optional starter labels. Without
them it returns the existing Chinese starter unchanged, so the other
callers (tests, desktop sidecar checks) keep their output.
- The web client passes translated labels from new diagram.seed keys in
en-US, zh-CN, ja and pl.
- fitDefaultDiagramLabels sizes the nodes and lays the starter out again
only when a translated label needs more room than the default label it
replaces, so the Chinese starter keeps its tuned geometry. The helper is
loaded with the lazily loaded diagram editor, which keeps dagre out of
the workspace chunk; if that import fails the note is still created
with the starter geometry.
WorkBuddy 5.8+ ships only dist/codebuddy-lite-wb.mjs. Its loader picks that
bundle only with CODEBUDDY_FORCE_LITE_WB_BUNDLE=1, otherwise --acp requires the
removed dist/codebuddy and exits. A bare --acp in the lite-wb bundle also runs a
one-shot print session; stdio ACP is served only together with --serve.
When the lite-wb layout is detected, set the bundle variable and launch with
--acp --serve --acp-transport stdio. --serve also starts WorkBuddy's local
gateway, so pin it to loopback with --host 127.0.0.1 and SERVER__HOST, which
overrides an inherited SERVER__HOST; the port stays WorkBuddy's random default.
Older layouts keep the previous launch, and the login flow adds only --port
when the command already serves.
Fixes#584
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
The packaged app is named @edgeever/desktop, so Electron stores userData at
<appData>/@edgeever/desktop. The reset guard only accepted a direct child of
appData and therefore always failed with unsafe-data-directory. Also accept a
single @scope/name directory while still rejecting other nested paths.
Fixes#583
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
* feat: add table context menu with copy and paste
* fix(web): preserve native input and test table context menu regressions
* fix: harden table context actions and desktop clipboard
Track close separately from exit, preserve probe outcomes on cleanup errors, and cover inherited pipes and Windows cwd locks. Add cross-platform CI with the official Codex ACP process chain smoke check.
Refs #523
Turn on both compiler checks in the root TypeScript configuration, which
apps/web, apps/api, apps/extension, packages/shared and packages/plugin-api
extend, and in the separate Expo configuration of apps/mobile.
No switch statement falls through today. The 19 members that override a
base class member (React error boundaries, the local database recovery
error and the vendored poster badge) now declare `override`; this only
changes type checking and is erased from the emitted JavaScript.
test:desktop listed 40 of the 49 Electron main-process test files by hand,
and the five ACP tests that need the desktop working directory ran in a
second command. Four files were missing from both lists
(desktop-device-model, edit-context-menu, rpc, window-lifecycle), so the
desktop build never ran them.
The first command now runs everything under apps/desktop/src/main and
ignores only the five tests the second command runs with --cwd
apps/desktop. The renderer tests and the sidecar step are unchanged, and a
new main-process test is picked up without editing the script.
* feat(memos): duplicate a note with its files
Add "Duplicate note" to the note list's context menu and the text editor's
"More note actions" menu. The copy keeps the note's content and tags, is
titled "<title> (copy)" in the UI language, and opens in the editor.
POST /api/v1/memos/:id/duplicate (write:memos) copies every image and
attachment of the note to a new resource with its own stored object and
rewrites the content to reference the copies, so editing or deleting either
note never affects the other. Shares, revisions, the pinned state, public
form links and unsubmitted form uploads are not copied. A note with more
than 20 files or more than 64 MiB of files is refused, and files written for
a copy that fails are removed.
The Web and desktop clients synchronize the note's queued changes before
duplicating and refuse while any remain, including unresolved conflicts;
the editor item stays disabled until the current edit is saved.
* fix(memos): keep committed duplicate files and check the whole desktop outbox
- Copied objects are removed only when the write batch fails. A failure
after the commit, such as a transient error while reading the new memo,
no longer deletes files that committed resource rows reference. When the
memo row is missing after the commit (its notebook was removed meanwhile),
the rows are removed before their objects.
- The sidecar lists at most 200 outbox items, so a conflict of the note
further down the queue was invisible and the stale server version could be
copied. The desktop client now compares the listed items with the
sidecar's status counts and refuses to duplicate unless the whole
outstanding queue was listed, nothing is being synchronized, and no item
belongs to the note.
- Regression tests cover a read failure after the commit and a target
conflict at position 201 behind 200 other conflicts.
Search raw Markdown in source mode with literal character matching, batched undoable replacement, and matching highlights. Preserve query whitespace and keep replacement focus stable. Route source editor search shortcuts through the workspace controller.
Validation: 141 editor tests and 2 Chrome end-to-end cases passed; web and mobile typechecks and the web build passed.
Refs #574
* test(api): cover public table form access, uploads and submissions
Run the public form routes and the owner form settings routes on SQLite
with every migration and filesystem storage: unknown, closed and
deleted-note tokens, password unlock and its rate limit, the fields a
form publishes, stale pending upload cleanup, field validation, and
which uploads a submission may attach.
* fix(api): accept file uploads on public table forms
Public form routes run without a signed-in user, but storing an upload read
the workspace from the auth context, so every valid upload to
POST /api/public/forms/:token/resources failed with a TypeError (HTTP 500).
In the form page the visitor's row was then saved without its file.
- Resource creation accepts the note's workspace explicitly, and the form
upload passes the form's workspace.
- A form accepts uploads only when it publishes an attachment field, and
holds at most 50 uploads that were not submitted yet (stale ones are
still retired after an hour). The limit is checked again after storing,
so concurrent uploads cannot exceed it.
- A submission marks as used only the files its new row references; a file
sent in another cell stays pending and is retired like other stale
uploads instead of being kept forever.
* test(api): cover JSON backup restore end to end
Export a workspace through GET /api/v1/backups/json and restore it into
another installation through the restore routes, on SQLite with every
migration and filesystem storage, the way the ZIP backup does:
- notebooks, notes, revisions, tags, pinned state and files round-trip
unchanged, and the restored notes are searchable and tagged;
- restoring the same backup twice changes nothing and keeps one object
per file, and re-restoring a file replaces and removes its old object;
- restoring brings back notes and notebooks deleted after the backup and
undoes later edits;
- damaged or minimal content falls back safely (empty document, Markdown
derived from JSON, default title, normalized tags);
- ids used by another workspace, notebooks or notes pointing outside the
workspace, revisions of another note, invalid file metadata, missing
notes, empty files and another user's file id are rejected without
writing;
- only signed-in users can restore.
Raises line coverage of backup-service.ts from 6% to 99% and of
backup-routes.ts from 46% to 94%.
* fix(api): validate a restore page before writing any note
restoreJsonMemos checked that each revision belongs to its note inside the
write loop, so a page whose second note carried another note's revision
returned 400 after the first note had already been restored. The check now
runs for the whole page before anything is written.
* test(api): cover note trash, restore and permanent deletion
* test(api): cover user management and login session endpoints
Exercise the user and session routes against an in-memory SQLite database with every migration applied and sessions created through real password logins: owner-only access, member provisioning, username and password validation, password resets and account disabling revoking sessions, per-device session listing, renaming and revocation, and per-user storage usage.
* test(api): cover object storage settings and credential handling
Exercise the object storage routes and resolution helpers against a fully migrated in-memory SQLite database with a stubbed S3 endpoint: owner-only access, secret redaction and encryption at rest, input validation, switching providers, per-resource storage resolution, fallback-key credential migration, undecryptable credentials and grouped deletion.
The full non-E2E suite (`bun run test`) only ran inside the release
workflow, and pull requests only built the Docker image, so a failing test
or type error surfaced at release time. A new Test workflow runs on every
pull request and push to main (Markdown and docs-only changes excluded):
- "Type check and build": `bun run typecheck`, `bun run typecheck:mobile`
and `bun run build:web`;
- "Regression tests": `bun run test` (bulk and integration suites).
It uses the Bun version pinned in package.json, retries `bun install` like
the other workflows, cancels superseded runs, reads the repository only,
and is limited to this repository so deployment forks spend no runners.
* fix(api): store LF line breaks for multipart image notes
multipart/form-data encodes line breaks in text fields as CRLF. Browsers
always do so, and Bun 1.4 now does as well, so notes created through
POST /api/v1/memos/with-image (used by the Web Clipper) were stored with
CRLF line breaks, while every other note uses LF. The route now normalizes
the Markdown field to LF before creating the note.
This also fixes the existing route contract test under Bun 1.4, and a new
test sends a CRLF multipart body like a browser does.
* test(desktop): make ACP host tests independent of the machine and Bun version
- The ACP command allow-list test checked that a missing command is not
installed, but on macOS the resolver also looks in /opt/homebrew/bin and
/usr/local/bin, so it failed on any machine with that command installed.
The test now only sees its temporary directory.
- The fake WorkBuddy agent started its account API without a port during
probes and chats. Bun 1.3 then listened on a random port, while Bun 1.4,
like Node, rejects port NaN, so six WorkBuddy session tests hung until
their timeout. Like WorkBuddy itself, the fake agent now serves the
account API only when started with a port.
tianma-if