queryEvents used to ignore its after/before arguments and ask Stalwart
for every event in the account with limit:1000. When an account holds
more than one page of events, the response was silently capped and the
grid rendered empty for the visible month — the missing events sat past
the truncation, never fetched. Meanwhile any recurring master whose
base start predates the window was returned exactly the same way, so
the client-side recurrence expander (already wired) never saw it.
Now the query sends after/before as a real filter ANDed with the
inCalendar restriction, and iterates position by page size until a
short page comes back (or 20 pages, as a bug backstop). Stalwart's
matcher accepts a master when any occurrence overlaps the window, so
the expander gets the raw event and produces the visible occurrences.
The refactor factors buildEventFilter() so the calendar / no-calendar
and window / no-window combinations produce the right filter shape
(single condition, AND, or absent). Tests updated to lock the new
filter shape and to cover the paging path (regression: 1000 events
that used to become 1000, now 1002).
Picked up from Jim Preston's uncommitted diff in the sister worktree
during the 2026-08-03 diagnostic session (artifact PkOWkT8C).
Signing in from the iOS build against a Stalwart server fails with "No
account found in JMAP session".
Stalwart's discovery endpoint /.well-known/jmap answers with a 307 to
/jmap/session. iOS NSURLSession strips the Authorization header when it
auto-follows that redirect, so the app gets an unauthenticated session
back — HTTP 200, empty accounts — and resolveAccountId throws. Browsers
keep the auth header on same-origin redirects, which is why the webmail
never saw this; it only bites the native app on iOS.
Request /jmap/session directly so the header stays attached, falling
back to /.well-known/jmap on 404 for servers that don't expose the
Stalwart path.
Also treat a 200 with no accounts as an auth failure. Stalwart answers
that way for missing or invalid credentials instead of sending a 401, so
a user who fat-fingers their password currently gets "No account found
in JMAP session" rather than "Invalid credentials".
Tests updated for both behaviours; full suite passes (the pre-existing
auth-store.test.ts parse failure on main is unrelated).
Two fixes App Review and testers hit on the iOS build of this app.
**Camera pre-prompt.** App Review rejects a custom pre-permission screen
whose button says "Allow" under guideline 5.1.1(iv) — it reads as an
imitation of the system dialog. Verbatim from the rejection: "A custom
message appears before the permission request, and to proceed users press
an 'Allow' button. Use words like 'Continue' or 'Next' on the button
instead."
The QR scanner's permission screen now says "Continue", and the copy
above it explains what the camera is for instead of asking for access.
When the OS stops allowing a re-prompt (canAskAgain is false) the button
becomes "Open Settings" and calls Linking.openSettings(), which is the
recovery path the same guideline suggests.
**Drawer first open.** The sidebar drawers park at -screenWidth and start
the slide-in from the effect that watches `visible`. That effect runs
before the Modal's native view is attached, so the very first open loses
its animation and the drawer renders almost entirely off-screen; opening
it again works because the view now exists. The open animation also runs
from the Modal's onShow, which commits the final offset once the modal is
really on screen and is a no-op on later opens. Applied to the mail,
calendar and contacts drawers, which all share the pattern.
Keeps the fix — gesture handlers now read current props — but makes the
class of bug structurally impossible rather than patched case by case.
The PanResponder is created once, so its handlers closed over the props of
the render that built it. Four call sites had to be corrected individually,
and nothing stopped a fifth from being added stale. Now the decisions live
in swipe-gesture.ts as pure functions taking the config as an argument, so
there is nothing left for a handler to capture; the component keeps one
`latest` ref pointing at the current props and hands it over on each event.
Also:
- Replace the placeholder test, which asserted that assigning to a locally
declared `{current}` object works and passed against the unfixed
component, with 20 tests over the real gesture logic: claim/decline,
overshoot clamping, commit thresholds, flick velocity, reveal open/close
from an offset base, and the 'none' cases.
- Build the responder in a useMemo instead of re-running PanResponder.create
on every render of every visible row and discarding the result.
- Fold the four repeated Animated.spring calls into springTo/settle.
Beyond the reported symptom, the same staleness broke two more cases:
entering selection mode passes 'none' for both actions, but a swipe still
fired the old action; and onAction captured a handleSwipeAction from before
the mailbox ids resolved, so on a slow connection swipe silently did nothing
without anyone touching the setting — which matches the deeper problem
Arnaud suspected on the PR.
Follow-up to the direct-view fix. No behaviour change on the happy path,
but the handoff is now easier to reason about and covered by tests.
- Use `File.contentUri` from the expo-file-system 19 API instead of the
legacy `getContentUriAsync`, dropping the `expo-file-system/legacy`
import. Same FileProvider, one less round trip through the bridge.
- Pull the intent launch into `openWithViewer`, which returns false when
the handoff didn't happen, so the fallback reads as a decision rather
than as an empty catch. It also no longer swallows the reason silently.
- Rewrite the expo-sharing comment above `shareAttachment`: it warns
against pre-translating to a content URI, which is still true for every
`Sharing.shareAsync` call but is exactly what the new VIEW path has to
do. Note why `downloadAttachment` keeps the share sheet as well.
- Record why FLAG_ACTIVITY_NEW_TASK is deliberately absent: the module
launches via startActivityForResult, which a new task would cancel.
Replying put `email.id` — the server-local JMAP handle, e.g. `gos2aabtux` —
into In-Reply-To/References. No recipient can match that against any
Message-ID, so the conversation splits in two: the original sits alone in
its thread while the reply (and everything answering it) starts a fresh one.
Proven against a live Stalwart account: the reply carried
`In-Reply-To: gos2aabtux` and landed in thread `btuz` while its parent stayed
in `btux` — Stalwart hashes the referenced msg-ids at ingest and found no
match.
Fetch `messageId`/`inReplyTo`/`references` with the full email and build the
headers from them (bracketed, ancestors first, parent closing the chain).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FTRyrFCR2N9qkpKK37GgWb
The dot added in #27 is positioned with `top: '50%'`, so it centres on the
whole row. Rows are three lines tall (sender / subject / preview), so the dot
lands a full line below the checkbox and the avatar it reads as a column
with — the taller the row, the worse it looks.
Anchor it on the first line instead: row padding + half an avatar, following
the density (and the text line height when the avatar is hidden in
extra-compact). The webmail's list carries the same `top-1/2` and the same
drift.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FTRyrFCR2N9qkpKK37GgWb
Fixes#32.
On Android, Sharing.shareAsync triggers Intent.ACTION_SEND (Share sheet). Updated shareAttachment to convert the file URI to a content URI and launch android.intent.action.VIEW via expo-intent-launcher, with fallback to shareAsync if no viewer application is installed for the MIME type.
Fixes#30.
PanResponder in SwipeableRow was created once on mount and captured initial props in a stale closure. Using mutable refs ensures gesture handlers always read and execute the latest leftAction, rightAction, mode, and onAction props.
Stalwart exposes task lists (supported-calendar-component-set = VTODO)
as sibling CalDAV collections, so Calendar/get returns them mixed in
with event calendars and the drawer showed confusing duplicate entries
(Todoist Inbox, per-project task lists colliding with same-named
event calendars).
JMAP Calendar/get doesn't surface the CalDAV component set, so the
store now infers the collection kind from its contents: a calendar
whose objects are all Tasks (>=1 task, no events) is flagged as a task
list. Flagged calendars are excluded from the Calendars drawer, the
event-creation calendar picker, and the ICS import target picker, and
sorted first in the Tasks sheet so new tasks default into a dedicated
task list. Empty calendars and mixed event+task calendars are
untouched.
The only unread cues were a font-weight bump and a grey-to-white shift
on the sender name, which barely render on some device fonts and are
nearly invisible in the dark theme. Mirror the webmail's dedicated
indicator (email-list-item.tsx): an 8px circle in the theme's existing
unread color, vertically centered in the row's start gutter, shown in
both the mailbox list and the unified inbox.
Verified on a Pixel 7 API 34 emulator in dark theme: unread rows are
unambiguous at a glance, read rows unchanged.
PR #20 (c5e5092) fixed#11 on Android by moving the editor page to
android/app/src/main/assets/editor.html, but the real bug was never
'webview scripting broken in release builds': the template literal in
buildEditorHtml cooked the insertLink scheme-regex escapes into an
invalid regex, and that SyntaxError killed the whole inline script.
The PR's buildEditorScript (injectedJavaScript) still contains the
cooked regex, so its theme/init injection dies on Android and the
editor stays completely dead on iOS, whose page no longer has any
inline script.
Resolve by keeping the lib-based editor from the local fix (valid,
escape-free inline script for both platforms, regression-tested in
src/lib/__tests__/editor-html.test.ts, verified on a Pixel 7 API 34
release build) and dropping the now-unused editor.html asset. All
other upstream fixes are kept.
The template literal in buildEditorHtml cooked the backslash escapes in
the insertLink scheme regex (\/ and \? became bare / and ?), so the page
received /^(https?:|mailto:|tel:|sms:|/|?|#)/i — an invalid regex. The
resulting SyntaxError killed the entire inline editor script: no change
messages ever reached RN, bodyHtml stayed empty, canSend never enabled
without an attachment, and mails sent with one had empty bodies.
Regressed in 29b06a3.
Rewrite the scheme check without backslash escapes and move the page
builder into src/lib/editor-html.ts so a unit test can assert the emitted
script parses and stays backslash-free.
Verified on a Pixel 7 API 34 emulator (release build): typing now posts
change messages and the plain-text body extraction sees the content.
Linus Rath
Linus Rath
lucletoffe
Ninad Chandorkar
arnauda-gh