| Age | Commit message (Collapse) | Author | Files | Lines |
|---|
|
Two defects, both reachable only after item 93.
startupSavedQuery() searched the saved queries alone, so a startup_query of
"Inbox" matched nothing once Inbox shipped as a built-in filter and the
duplicated entry was removed from queries.json. It then fell back to
m_savedQueries.first(), which is an arbitrary choice that used to look
reasonable while every install carried an Inbox entry: with the duplicates gone
it opened on a leftover search for one sender, and an empty queries.json opened
on nothing at all. The search now covers the saved queries first, so the user's
own entry wins a name collision, then the built-in filters; the fallback is the
Unread filter, which is always present.
The default startup name has always been "Unread" and now resolves for the
first time: before this it named nothing unless the user happened to have such
an entry.
The constructor also read startup.query directly, and a generated entry stores
no query at all, so even a matching filter opened an empty bar. It resolves
through Config::resolvedQuery() now, unscoped, since the account dropdown
starts on "All accounts".
Icons per the user's choices: a star for Important rather than the flag
action's own icon, since on the query row an icon reads as a category rather
than as an instruction, and mail-folder-sent rather than mail-sent.
Three tests changed rather than adapted, because their premises were the
defect. Two asserted the first-saved-query fallback.
aCronSyncDoesNotRefreshBeforeAnyQueryHasRun assumed a fresh window had run no
query, which is no longer true; it is now
aCronSyncRefreshesTheLastRunQueryNotTheQueryBar and asserts the property that
actually matters on a cron timer, through a new lastRunQueryForTesting() seam,
since a legitimate refresh bumps the generation and the counter cannot tell the
two apart.
|
|
Item 93, the UI half. Unread, Inbox, Flagged and Sent are buttons the
application ships, sitting first on the row, ahead of the user's pinned saved
queries.
runFilter() is runSavedQuery()'s opposite in the one way that matters: it READS
the account box and never writes it. That is item 90's defect. A filter narrows
what the user is already looking at, so the dropdown is its input rather than
something it resets on the way past. A saved query keeps setting the account
from what it stored, because it is a destination and states its own scope.
runQuery() gains an AccountScope parameter. A filter's text arrives already
resolved in the selected account's scope, and scoping it again would put
path:"work/Sent/**" inside path:"work/**".
Two migration changes, both of which unpin rather than delete:
- Sent is no longer migrated from the INI into queries.json. The built-in
filter covers it, and migrating one too would put two Sent buttons on the
row, one editable and one not.
- A stored entry naming a known generator is unpinned on load, which is what
every install upgraded through 0.19.0 carries. It keeps its name and its
generator and moves to the menu. Deleting it would be data loss on a file
whose readers are supposed to preserve what they do not own.
The test suite needed the same distinction the design makes.
savedQueryButtonLabels() now skips the filters, and savedQueryButton(window,
label) replaces five positional row->findChild<QPushButton *>() lookups that
were silently returning Unread.
One rendering probe had to be fixed rather than adapted.
replyRowsKeepTheirTextUnderTheThreadLine resized the window to 300px, and four
more buttons pushed the reply row below the viewport: the pixel loop then ran
zero times and reported "0 pixels, the row was painted over", which is a
different defect from the one it exists to catch. It gets 600px and a guard
asserting the row is really inside the viewport, so the next person to shrink
it gets told the truth. Verified by putting 300 back: the guard names the row
at 83..165 in an 82px viewport.
|
|
Item 74. runQuery() set the status bar once and only queryFinished cleared
it, so the bar kept claiming a query was running for the whole walk while
rows were visibly arriving behind it. Measured cold against a 1.1 GB index:
the first batch reaches the model at 642 ms and the walk finishes at 5714 ms,
so five seconds of a slow query read as a frozen one.
onThreadsReady now sets the bar from the model's own row count after each
batch, which is the number of rows the user can actually see. No timing
changes; this only stops the bar from lying.
The refresh branch returns before the new line, so a background refresh stays
silent exactly as onQueryFinished already keeps it. That silence has its own
test, which fails when the write is moved above the guard.
beginRefreshForTesting() is a new seam: refreshCurrentQuery() returns early
without a worker and a bare window has none, so a test cannot otherwise reach
the refresh path.
|
|
Selecting a thread root used to render the whole conversation, stubs
plus the last messages expanded, but only until the thread had been
expanded once. After that the identical click rendered a single message.
The user reported the inconsistency and asked for the single-message
behaviour throughout, and for the conversation view to go.
The cause was a timing one, not a race. The root card stands for the
thread's first message and onThreadSelected already preferred to load
just that, but the model learned the id only when the replies arrived,
so a fresh row fell through to a whole-thread render.
ThreadSummary now carries firstMessageId from the query itself, so the
id is known before any expansion and the fallback is unreachable. It is
free: notmuch_thread_get_toplevel_messages reads the index, not the
message files, and a walk with it is indistinguishable from one without
over a 36,615-thread database. Contrast recipients, which reads every
file and stays Sent-only.
The Sent view keeps showing what the user sent rather than the thread's
opening message, which is often someone else's. There is no
matched-messages iterator in libnotmuch, only a count, so that branch
walks oldest-first to the first NOTMUCH_MESSAGE_FLAG_MATCH and stops:
0.146s against a 0.143s baseline over 4,515 threads.
onThreadLoaded merges into renderMessages, since onMessageLoaded was
already delegating to it for the actual painting. It still takes a list
because MessageView renders a list; collapsing that is a separate change
to a class with its own tests.
NotmuchWorker::loadThread is kept and documented as having no UI caller.
It is a tested way to read a thread's messages with the match set
resolved, used as a helper by the worker's own tests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
showWarnings() did two separable things and one of them could not be
reached from a test. It set the status label, which is harmless, and it
raised a QMessageBox from the MainWindow constructor, which under the
offscreen platform nothing can dismiss: the constructor never returned
and the suite hung with no output, reading as an infrastructure failure
rather than a test one.
It splits in two. applyWarnings() keeps the status label and stays in
the constructor. configProblems() returns the list, and main.cpp raises
the dialog after show(), which also gives it a visible parent to sit on.
The distinction between warnings and problems is preserved exactly: a
keybinding being ignored interrupts startup, "no sync command
configured" does not.
The warning path now has its first test, using the config shape that
caused the original hang. Mutation checked by putting the modal back in
the constructor: the test times out at 124 rather than failing, which is
the behaviour this removes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Four signatures, no behaviour change: the two shipped operations map to
Replace and Narrow. runSearchFromPane becomes a switch and gains the
Exclude arm, which nothing can reach until the menu entry exists.
Seven call sites across three test files moved with it, two more than
the plan predicted: test_messageview and test_mainwindow also drive
these signals directly.
mainwindow.h and messagedetailsdialog.h now include searchterm.h for
the type; messageview.h already did.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The panes carry a finished query and know nothing of the query bar; the window
sets the field and calls the existing runner, so the account scope and the
generation counter keep working as they do for a typed query.
Narrowing combines here rather than in a pane, because only the window can see
what the bar currently holds. The tag strip's chips join the header, the body
selection and the details dialog as a fourth surface.
Also fixes the details dialog to actually close when a search is chosen: the
comment above the connection already described this requirement, but nothing
called accept() or reject(), so the dialog stayed open, the query ran behind
it, and the modal exec() never returned. This hung the whole test suite on
QT_QPA_PLATFORM=offscreen once a covering test was added.
|
|
The dialog is non-modal and single-instance, so a second Create tagging
rule reaches one that is already up. Seeding it beats dropping the
request, which would read as a broken menu item.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Item 82. Saving a query worked and nothing else did: changing one field meant
retyping the whole query under the same name, and deleting one meant editing
the file by hand. An action that creates something the UI cannot then change or
remove is incomplete, and the user hit it within minutes of the first hand test.
Right-clicking a saved query, on its button or its menu entry, now offers Edit,
Move to menu / Show as a button, and Delete. Every path funnels through one
replaceSavedQuery(), which matches on the name the dialog was OPENED with rather
than the one it returns, so a rename replaces the entry instead of leaving the
original behind beside a new one, and which merges the stored entry's unknown
fields in a single place rather than in three. Delete confirms first: the rule
against confirmation dialogs covers tag mutations, which the undo stack can take
back, and this writes user config that it cannot.
Two cases the item did not anticipate. A generated entry has no query to edit,
so the dialog shows its composed query read-only rather than offering a field
that changes nothing, and carries `generated` and `flat` through an edit rather
than letting it decay into a plain entry holding a snapshot of what it resolved
to today. And the overwrite notice had to learn to ignore the entry being
edited, since warning that "Inbox" already exists while editing Inbox is noise.
This also fixes a defect that predated it and was already reachable from the
save path. rebuildSavedQueryRow() called deleteLater() on the old row, which
defers destruction to the event loop, so the stale row went on answering
findChild() and every lookup after a rebuild reported the state from before the
edit. Nothing looked wrong on screen, which is why it surfaced only as three
tests failing against a row that had in fact been rebuilt correctly.
Five tests, three mutations. Matching on the returned name fails two, never
writing the file fails three, and dropping the unknown-field merge fails one.
That last one initially proved nothing: it drove UNPIN, which copies the stored
entry and so carries `unknown` along by itself, and passed with the merge
deleted. It now goes through the edit path with a replacement that has none,
which is what the dialog actually returns.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The spec asked for "a Save query button beside the search bar" and what shipped
was a menu entry and Ctrl+S. The user went looking for the button where the
design said it would be and did not find it.
Saving is a thing you decide on while looking at the results, so it belongs
where the results came from rather than behind a menu or a remembered chord.
The button takes the action through setDefaultAction rather than a second
connect, so it inherits the text, icon, tooltip and enabled state and cannot end
up offering to save an empty query while the menu entry correctly refuses. The
mutation that replaces it with a plain clicked() connection fails the test.
Also records item 82: a saved query cannot be edited, unpinned or deleted from
the UI. Item 23 specified saving and nothing else, and that is exactly what was
built, so the only way to unpin a query is a text editor or retyping it in full
under the same name. An action that creates something the UI cannot then change
or remove is incomplete, and this was found within minutes of the first hand
test. It is filed as a defect rather than an enhancement, and the spec now says
so where a reader would otherwise take the design for complete.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Second half of item 23, on top of the storage change. A query can now be kept
without hand-editing a file, and the row of buttons no longer grows without
bound.
Ctrl+S opens a dialog on whatever is in the query bar, taking a name, an
optional account scope and whether the query is pinned. It preselects the
account already chosen in the dropdown, since that is the scope the user is
looking at, and it says so when a name is about to replace an existing query
rather than refusing the name: overwriting a saved query on purpose is a normal
edit, and the only thing worth preventing is doing it without noticing. Saving
over an entry keeps the stored entry's unknown fields rather than the dialog's
fresh value, so a field written by a later build survives being edited here.
The saved queries move to a row of their own beneath the query bar, pinned ones
as buttons and the rest behind a More queries menu that only exists when
something is in it. The ponytail note that stood in the query row predicted
exactly this: an unbounded list of buttons sharing the row squeezed the field.
Sent moves down with them and is still not a saved query, for the reason already
recorded there.
A saved query's account scope goes through the account DROPDOWN rather than
being baked into the query text. runQuery() already wraps the query in the
selected account's path, so pre-scoping here would apply it twice, and setting
the dropdown also shows the user which scope they are in. An unscoped query
clears the selection rather than inheriting whatever the last one left, which is
the same defect the rules preview had.
Seven tests, three mutations. Ignoring the pinned flag fails two of them,
pre-scoping the text instead of setting the dropdown fails two, and letting an
unscoped query inherit the previous account fails one. The menu-absence test
initially passed against no implementation at all, since it only asserted a
widget was missing; it now proves the row was populated first, which is the
guard that class of test needs.
Two existing invariants caught real omissions rather than needing adjustment:
every registered action must appear in KeyMap::knownActions(), which is what
gives it a configurable binding, and every action needs its own icon.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Item 77. The dialog could say how many messages a rule matched and not
which ones. A Preview in list button now runs the selected rule's query
in the main window; the dialog stays open, since comparing the rule
against its results is the point.
Two constraints from the backlog entry, both now asserted and both
mutation-checked.
The query runs exactly as stored, with no tag:new and no wrapping
parentheses. The post-new hook adds those when it applies a rule, and a
preview that copied them would match nothing outside a sync window,
since tag:new is set only on mail that has just arrived.
The account selector is cleared first. runQuery() wraps the bar's text
in the selected account's scope, and a rule query usually names its own
path already, so previewing one with an account selected would scope it
twice and show an empty list, which reads as "this rule collects no
mail".
The second mutation only fails once the test's config has an account to
select: with the default empty config the selector sits on "All
accounts" anyway, and asserting that a preview leaves it there passed
against the mutation. Recorded in the test.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Counts are generation-stamped and dropped when stale or when the dialog
has closed: counting every rule against a cold index takes seconds, so
an in-flight reply outliving its dialog is ordinary rather than rare.
The stamp is its own counter, not m_generation as drafted. That one is
the QUERY generation, compared against directly by every thread, tree
and message load, so bumping it to count rules would discard whatever
the user was opening at the time and blank the message pane for an
unrelated reason.
Registering an action obliges two more entries, both enforced by tests:
the name in KeyMap::knownActions(), and a default binding, since every
action carries one. Ctrl+Shift+T, shifted against Ctrl+T for edit_tags
the way Ctrl+Shift+U is shifted against Ctrl+U.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Item 71. A tag edit reached the notmuch index at edit time and then sat there
until the user clicked Sync or their cron job fired, so "mark all read" updated
the view while the change itself waited, sometimes for ten minutes.
A confirmed edit now arms a debounce that runs the existing sync path. The delay
is auto_sync_delay_ms in [general], defaulting to 2000, and follows
mark_read_delay_ms exactly, including that zero and negative are not errors:
zero syncs on the next trip through the event loop, and any negative value
disables the behaviour, which is the switch for a user who wants only their cron
job.
It is armed from onTagsApplied, where a write is confirmed and the pending count
is already current, rather than where one is sent: a sync scheduled for a write
the worker went on to reject would run for nothing. A debounce rather than a
schedule, restarted by each edit, because "mark all read" confirms one write per
thread in the view and an arm-per-edit timer would be the storm of syncs the
debounce exists to prevent. Nothing is armed when no sync command is configured
or when the pending count is zero, the case where an edit was netted against its
own inverse. When the timer fires with a sync already running, local or cron, it
skips rather than queues: mbsync's own answer to a second run is to fail on it,
and the edits stay pending rather than being lost.
Also fixes a pane blanked out from under the reader, found by hand testing this
feature. onSyncFinished called runCurrentQuery() where the cron path calls
refreshCurrentQuery(), and a re-run clears the model, the undo stack and the
message pane. The stale-thread notice handles a thread that stops matching the
query and has since item 35, but a re-run left nothing for it to describe. The
two paths had no reason to differ; before this item a local sync only followed a
click on Sync, so the difference went unnoticed. Reading a message in the Unread
view, having it marked read, and watching the pane go blank two seconds later is
what surfaced it.
Its test asserts on the undo stack rather than the pane: both paths issue a
queued query test_mainwindow has no worker to answer, so the pane ends up blank
either way and an assertion on it would pass against both, while the undo stack
is cleared by one and kept by the other.
Nine tests, four in test_config and five in test_mainwindow, each
mutation-checked: removing the schedule call, honouring a negative delay,
dropping the nothing-pending guard, dropping the already-running guard, and
restoring runCurrentQuery() each fail a test.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Item 67. The pane counted unread, flagged and inbox from three fixed
tag: queries. Sent and drafts cannot join that list as tags: tag:draft
counts 0 against a real database and no draft-ish tag exists in it at
all, so a tag-based line would be a permanent zero that reads as working
code. Both are composed from each account's folder keys instead, the
same way the Sent view already composes its query.
The drafts key was parsed and documented as unused in v1. Composing
drafts is still v2; counting them is not, so Account::draftsQuery() and
Config::allDraftsQuery() now mirror the sent pair. The shared body moved
into folderQuery() and joinAccountQueries(), so the load-bearing quoting
(a provider nests both folders under a bracketed parent, and [ and ] are
Xapian syntax) and the bare-"or" guard exist once rather than once per
folder type.
The fixed array is gone rather than extended. It held queries and labels
in two lists indexed in parallel, which is a hazard that grows with the
list: an entry inserted in one and not the other prints a real number
against the wrong name and looks entirely plausible. placeholderLines()
carries each query beside the callable that labels it, so the two cannot
drift, and the count reply stays paired by position as the worker
requires.
A line is omitted when no account configures that folder rather than
shown as 0, following item 63: a missing folder is a real configuration,
and "0 sent" claims the user has sent nothing.
Measured against the real config: 4 sent terms over 601 threads, 5
drafts terms over 3, the extra drafts term coming from the one account
that configures drafts and no sent, which is what proves the two are
collected independently.
Four tests here and four in test_config, mutation-checked at three
points: dropping the drafts line, an off-by-one in the label pairing,
and removing the -1 guard for an uncountable query. Each mutation fails
a test.
|
|
Adds a `sent` key to [account.*] naming that account's sent folder, and a
Sent button beside the saved queries that composes its query from every
account carrying one. An account without the key is omitted silently, as a
real account may keep no sent mail locally. With no account selected the
button spans all of them; selecting one narrows it through the existing
scope wrap rather than a second path.
Composed at run time rather than shipped as a [queries] entry. A saved query
is one fixed string: it cannot narrow to the selected account, and it goes
stale the moment an account is added or a provider renames a folder.
The design and the measurements behind it are in
docs/superpowers/specs/2026-08-11-sent-mail-design.md. Three things there are
worth repeating here.
The composed path is QUOTED, and that is load-bearing. A real provider nests
its sent folder under a bracketed parent, and "[" and "]" are Xapian syntax:
unquoted, the query parses rather than matches and returns nothing while
looking entirely plausible. Composition happens in one place so there is one
chance to get it right, and a bracketed path is pinned in a test.
Recipients are opt-in per query, which is a performance contract rather than
a preference. notmuch_message_get_header(m, "To") is not served from the
index, it reads the message file: folding every thread of a 4411-thread
inbox took 38.2 seconds against 251 ms for the 601-thread sent view. The
worker skips the walk entirely unless asked, and the refresh path carries the
same flag so a background sync cannot blank the column mid-read. Always
folding is mutation-tested: the data would be right and only the cost wrong,
which nothing else here would notice.
The messages reached through the thread are owned by it and freed with it, so
recipientsOf() holds them raw and finishes while the thread is alive, exactly
as walkReplies does. An NmMessage wrapper there is a double-free.
Sent mail is presented flat, and the pane follows. A message you sent
otherwise drags in the replies you received, so a view labelled Sent shows
conversations rather than what you sent. ThreadListModel::setFlatMode() makes
hasChildren() and ReplyCountRole answer differently and changes nothing else;
runQuery() sets it on EVERY run, so any other query restores the tree on its
way through and the flag cannot outlive the button that set it. The pane
needed its own fix for the same reason: the single-message path depends on a
field only filled when a thread is expanded, which never happens in a flat
list, so loadThread() gained matchedOnly and drops the messages that did not
match instead of rendering them as stubs.
Recipients replace the sender through the existing SendersRole rather than a
new one, so the delegate needs no branch and cannot disagree with the model
about which name a row shows. It falls back to the sender when a To header is
absent or unparseable, since a blank where a name belongs reads as a
rendering fault.
Address parsing uses GMime: a display name may contain a comma, so
"Rossi, Mario" <m@example.org>, info@example.net is two addresses and
splitting reports three. internet_address_list_parse returns NULL for an
empty string, which is a crash if unguarded.
Backlog item 63.
|
|
The thread list now updates itself when a sync finishes, whether it is
empty or populated. New threads appear where the sort puts them, threads
that stopped matching leave, and threads whose state changed repaint.
Refreshing used to mean re-running the query, which cleared the model,
the selection, the message pane and the undo stack, so 0.8.0 declined to
do it on a cron timer and asked the user to press Enter instead. The
result was a list that quietly disagreed with the database: mail indexed
by cron never appeared, and an Unread view read to the end sat empty in
front of it.
ThreadListModel::reconcile() diffs a result against the current rows by
thread id instead, so a surviving thread keeps its row, its persistent
index and its loaded replies. Order comes from the result and is never
imposed here, which is what makes the sort dropdown authoritative.
The undo constraint this was sized around did not exist: no undo entry
was ever keyed on a row. ThreadTagCommand stores thread ids and
MessageTagCommand stores message ids, and applyTagChange() looks its
target up by id, so an entry already survived its rows leaving the view.
A thread read out of the current view now leaves the list, which is
correct and would otherwise strand the reader, so MessageView grows a
notice saying the open thread no longer matches, with a button that
re-queries it. Recovery lists the whole conversation, expands it, and
restores the message that was on screen rather than reopening at the
first one.
Ten defects were found building this, nine of them by hand testing:
- SyncMonitor::start() polls synchronously, so an idle lock file emits
stateChanged(Idle) from inside buildUi() and the first handler to
touch a widget segfaults before the window exists.
- QTreeView sets a current index when it takes focus with none set, and
current drives loading, so new mail opened itself and was marked read
without the user having looked at it. Selection is now required.
- The notice outlived what it described, both when the pane was blanked
and when another message replaced it.
- Retiring the "Background sync completed" message left the bar claiming
a sync was still running: silent means saying nothing new, not leaving
a stale claim on screen.
- A thread root sets both the thread id and the message id, so treating
the message id as the message-row case discarded it for the commonest
way to open a thread.
- A freshly queried root does not know its own first message until the
tree loads, so recovery selected nothing and left the pane blank.
- A user query mid-recovery had its result hijacked by the pending
selection.
- MessageView emitted the recovery signal with its own members, so a
direct connection handed MainWindow references that runCurrentQuery()
then cleared by blanking the pane. The ids went empty mid-slot and no
recovery ever ran. Every test passed against this, because reaching a
slot through invokeMethod copies its arguments.
A Qt signal argument is a reference until something copies it. Emitting
a member to a slot that can re-enter the emitter is a use-after-write,
and it presents as a wrong value rather than as a crash.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Two entries, straight to notmuch. This adds a feature rather than replacing one:
the column header was decorative and nothing implemented click-to-sort, so
removing the header with the grid lost nothing.
Stored in uistate.conf, never in the hand-edited config, and range-guarded on
read: a stale file can hold anything, which is the lesson item 58 recorded.
SortOrder needed qRegisterMetaType despite carrying Q_ENUM. Q_ENUM gives the
enum a meta-object entry, not a metatype registered under the name invokeMethod
resolves, so the queued runQuery would have dropped its sort argument at runtime
and every query would have silently run newest-first. Nothing in the suite
exercises a real worker thread, so this was asserted directly rather than left
to a warning nobody would see. It is registered beside the type rather than in
MainWindow's constructor: a first attempt put it there and passed only because
the test that catches it never constructs a MainWindow.
The account dropdown's entries now carry their account's colour as a swatch,
which is what makes the accent bar on a card mean anything: a colour down a
card's edge says nothing until something maps it to a name. Raw colour here
rather than the blended line colour, since a swatch is a filled patch like a
chip rather than a thin line. Its test builds its own two-account config: reading
the environment's made it SKIP wherever no accounts are configured, which is a
test that asserts nothing while reporting success.
|
|
tagSelected resolved rows to threads with threadAt(index.row()), which is wrong
for a message row: a child's row number indexes its siblings, so acting on a
reply tagged whichever thread sat at that position in the list. It now routes
through ThreadListModel::scopeFor, and a message row's change is sent as message
ids down applyTags with its own MessageTagCommand for undo.
MessageTagCommand stores message ids where ThreadTagCommand stores thread ids,
and that difference is the point rather than an inconsistency: re-resolving the
thread on undo would restore tags across every sibling the action never touched.
sendMessageTagChange deliberately skips the optimistic model update.
applyTagChange is keyed by thread and would repaint the whole row as though
every message in it had changed, which for a one-message edit is a lie the user
watches correct itself on the next query. It keeps the two things that are NOT
optional: the edited-account set, resolved through the containing thread since
the account is a property of the thread, and holding the edit when a sync holds
notmuch's write lock, since the worker's read-write open blocks rather than
failing.
The scope is now stated before an action and after it, naming both the message
count and whether a whole thread went. This is what stands in for the
confirmation dialog CLAUDE.md rules out: undo is the safety net, and undo is
only usable if the user can tell that something larger than they meant has just
happened. Selecting a single message reports no count at all, since reading one
message is not a bulk action.
A mutation that routed message rows down the thread path SURVIVED the whole
suite: undo depth and status text are identical either way while every sibling
gets tagged. anActionOnAMessageRowTagsThatMessageNotTheThread exists because
that gap was found, and asserts on the ids actually sent.
anActionOnAThreadRowSaysItHitTheWholeThread reads the status bar BEFORE draining
the event loop. This binary has no worker, backlog item 36, so the queued write
reaches a database that has never heard of the thread and answers with
errorOccurred, which overwrites the status bar: draining first asserts on that
error and fails against correct code.
|
|
loadMessage queries by id: and returns one MessageRef, always matched, since the
user asked for that message by clicking its row and a stub would answer the
wrong question. An unknown id emits an empty vector rather than an error: a
stale row after a reindex is an ordinary race, not a failure worth the status
bar. The signal fires even when empty so the UI handler runs instead of waiting
for a reply that never comes.
The branch in onThreadSelected is placed BEFORE threadAt(), which is the whole
trap. threadAt takes a top-level row number and a child's row number indexes its
siblings, so handing a message row's number to it loads whichever thread happens
to sit at that position. Mutation-checked: with the branch disabled the test
reports thread 't1' for a reply belonging to 't2', a wrong answer plausible
enough to survive review.
m_currentMessageId and m_currentThreadId are mutually exclusive and each clears
the other, so a queued reply can tell which kind of selection it belongs to.
onMessageLoaded carries a third guard onThreadLoaded does not need: a reply
landing after the selection moved to a thread row would render one message where
the conversation belongs.
No mark-read timer for a message row in this pass. Marking one message of a
thread read is a per-message tag write and the pending-edit map is keyed by
thread; item 28 is the record of what happens when that count goes wrong.
|
|
Replies are fetched on expansion rather than with the query: walking the reply
tree of every thread in a 10k-thread result would cost more than the query and
almost none of it would be looked at.
hasChildren is what makes that lazy loading work, and its absence would have
shipped the feature unreachable. rowCount is 0 until the worker has walked the
thread, so a view left to infer the expander from rowCount alone draws none, the
user can never expand, and the replies are never requested. It answers from the
summary's totalCount before loading and from the children afterwards, so a
thread whose count included duplicates stops offering an expander that opens
onto nothing.
onThreadTreeLoaded reads the thread id from the reply rather than remembering it
from the request. Two expansions can be in flight at once, and pairing them by
order would attach one thread's replies to the other.
|
|
The strip survived the port because every geometry call it needs exists on both
classes. What did not survive is anything keyed on a row NUMBER: a tree numbers
rows per parent, so row 0 exists once per expanded thread and the old flat 0..N
walk would paint the first thread's strip over every one of them. The walk now
goes by index, and the alternating colour follows visual position rather than
index.row() for the same reason.
QTableView::isRowSelected(int) has no QTreeView equivalent; isSelected on the
index replaces it. MainWindow loses verticalHeader and selectRow, so row height
comes from uniformRowHeights and three helpers replace the row arithmetic.
next_thread and prev_thread now resolve the containing thread first: in a tree
current.row() + 1 is the next SIBLING, which under an expanded thread is the
next reply, not the next thread.
Two test defects found by mutation and worth recording, since both produced a
green suite over a broken assertion:
The indent test asserted on column 0. A QTreeView indents only the column
holding the expander, verified against Qt 6.11: with setTreePosition(4), column
0 reports the same left edge for a thread and its reply while column 4 reports
420 against 440. It was failing against a correctly indented tree.
The strip test passed with the view's skip deleted, because the real model
already returns no pills for a child row, so the view's guard was never the
thing under test. It now runs against a stub model that hands pills to every
row, which leaves the view's skip as the only thing that can keep replies clean.
That rewrite then failed for a third reason: without the delegates MainWindow
installs, rows take the default height, the band is measured against
SubjectDelegate::rowHeightFor and overflows into the row below, and the thread's
own strip paints across the reply. Reads exactly like a missing skip and is not
one.
|
|
A splitter position is saved in pixels, so one saved in a wide window
does not fit a narrower one: QSplitter::restoreState() honours the first
pane's saved size verbatim and gives the second whatever is left. A real
1285/1252 split restored into a 1136px window left the message pane 29px
wide, a sliver of rendered mail beside a full-width thread list. The
wider the window ever was, the worse the next narrower session.
Fixed with a minimum width on the pane and setCollapsible(1, false),
which covers the restore and the equivalent drag. A restore-time repair
running from showEvent() was written first and deleted: with the floor
in place it was mutation-tested to be redundant, since the minimum width
constrains restoreState() as much as it constrains a drag.
The floor is 300px rather than a bare "visible" width, because it is
reached only when a position does not fit and should land somewhere mail
is readable; at 200 the placeholder's own text wraps every few words.
The test asserts against the pane's own minimumWidth() rather than a
repeated literal, with a guard on the floor itself so it cannot pass
against a lowered one.
Note for later work in test_mainwindow: the offscreen platform chooses
the window width itself and has been seen to choose differently between
two runs of the same binary (1181 and 779), and it ignores resize(),
setFixedWidth() and a resize of the splitter on a shown window. Any
assertion on a pane ratio, or on the second pane's pixels, is measuring
that choice rather than this code.
Backlog item 55, whose recorded cause blamed the thread view's size hint
and first-run layout. Measured, that hint is 256px, not the 886 the item
computed from the column widths, and a freshly built window splits
correctly; the entry has been corrected in place.
|
|
Item 54. A sync fired by the user's cron carries tag edits to the mail
store exactly as a local one does, but only onSyncFinished() cleared the
pending state, so the indicator kept reporting work that had already
shipped and the exit prompt asked to sync for it. Verified against a
real cron run: 31 changes, cleared with no manual sync.
The window cannot see an external run's exit status, and /proc/locks
carries no outcome. It does not need to: mailsync.sh already ends every
run with a "RUN END ... status=OK" banner in its log, which outlives the
process that wrote it. MailSync::lastRunOutcome() reads a bounded tail
of that file and takes the last completed marker, so no change to the
script and no optimistic guessing were needed.
Only a definite OK clears anything. A failed run, a missing or
unreadable log, and a State::Unknown lock reading all leave the count
alone: over-reporting costs a redundant sync, under-reporting costs the
user their edits.
m_editedAccounts is drained in the same place, before flushHeldEdits()
and matching the local path's ordering. Item 49 uses it to choose which
mbsync channels a run syncs, and a count that reached zero while the set
stayed full would look correct and still sync the wrong channels.
The log path comes from a new optional [sync] log key, defaulting to
where the script writes, so a test never reads the developer's own log.
Two notes on the verification, both recorded in the backlog:
- A timing probe endorsed a tail read that was not happening. The
first version of the huge-log test required the call under 100 ms
and passed with the seek deleted, because reading 10 MB is fast
either way. Replaced with an assertion on content.
- Every fixture was invented and the first batch had the wrong
timestamp format, since the script uses date -Iseconds. The tests
passed anyway, because the parser keys on the prefix and the status
token. One test now builds the banner the way the script does.
The before-flushHeldEdits ordering has no test: without a held lock the
flush is a no-op, so both orderings pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
A sync ran mbsync -a regardless of what changed, so tagging mail in one
account fetched all of them. The account set was not a parameter anywhere
on the path: MailSync::start() took no arguments and the script hardcoded
-a, so nothing between a tag edit and mbsync carried which account changed.
Track which accounts have edits and pass their mbsync channels through to
the script, which now takes channel names and falls back to -a when given
none. An empty set means all accounts, per the request: a sync with nothing
pending is a fetch, and narrowing that to wherever the last edit landed
would quietly stop collecting mail everywhere else.
The channel is a new optional per-account key rather than the section key.
The two names genuinely diverge, because a QSettings section key may carry
dots that the channel does not, and mbsync treats an unknown channel as
fatal rather than skipping it, so key-as-channel would fail those accounts'
syncs outright rather than degrade. It defaults to the key, so accounts
whose two names already agree need no config change.
The edited-account set is deliberately not netted the way the pending-edit
map is: that map tracks the index, where a tag removed and re-added leaves
nothing outstanding, while this tracks the mail store, where both writes
have already renamed files that mbsync still has to propagate. It is also
snapshotted before flushHeldEdits(), which inserts into it synchronously
rather than on a queued reply, so a successful sync cannot clear accounts
whose edits it never carried.
Closes item 49.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Nothing in the UI reported database-level facts: every query gave a
thread count for that query, and nothing said how much mail there is
overall. A dialog under Help now shows messages, threads and tags from
notmuch, plus the account list from config, since notmuch does not model
accounts at all.
A separate worker call rather than a reuse of requestCounts, which counts
threads to match the row count of a query. This counts messages, which is
what a user means by "how much mail is in here". The test pins 4 messages
in 3 threads against the fixture and fails if they are ever made equal,
so routing both through one count cannot pass unnoticed.
Every field starts at -1 and renders as "unknown" when notmuch could not
answer it. Printing 0 would say the Maildir is empty, and telling someone
their mail is gone is the worst way to report an index that failed to
open.
The dialog opens showing "Counting..." rather than blocking, since
counting every message is not free on a large database. That makes two
lifetimes matter: the reply can arrive after the dialog is closed, so the
label is a QPointer, and the dialog can be closed and reopened while a
count runs, so a generation counter drops the older answer. The test
drains DeferredDelete before firing the late reply, because close()
deletes through deleteLater and without that the dangling case is never
actually reached.
|
|
An empty right pane said nothing, and multi-select made it a routine
sight. It now carries the wordmark, thread counts that run their query
when clicked, and a sync line that appears only when something needs
attention.
Rendered into the existing web view as a third document shape, so there
is one document path and one set of security rules. The brand palette is
a deliberate exception to deriving colours from the desktop theme, since
a logo is brand rather than chrome; the theme still picks which of the
two sets is used.
Counts refresh when the pane is about to show rather than in the
background: one goes stale the moment a tag is edited, and refreshing one
nobody is looking at is work for nothing. A generation counter discards a
superseded reply, and a late answer cannot repaint over an opened thread.
The helper lines are real links because JavaScript is off in this
profile. The handler is gated on the placeholder actually being
displayed, so the same URL inside a message body is dropped: a stranger's
mail must not drive the thread list, even to run a harmless query.
Three defects found while building, all silent:
- Every CSS percentage was invalid. QString::arg does not collapse "%%"
into "%", so the document carried "50%%" and the browser dropped each
declaration holding one, disabling the mask, the glow and both radial
gradients while still rendering something plausible. Substitution is by
named token now, which cannot collide with a percent sign.
- A geometry probe endorsed the layout while that was live, because it
measured only properties without percentages.
- The font test passed against a build with one face missing, since the
other satisfied both of its checks on its own.
The mockup's light values needed correcting against a real pane: the grid
vanished at a 2% luminance step on white, and the glow subtracts light
there rather than adding it, washing the pane. Strength only, not hue.
|
|
An action removing "unread" from every thread in the current view,
on the toolbar, the Message menu and Ctrl+Shift+U. It deliberately
ignores the selection, which makes it the one action in the window
that does, and it routes through the same funnel as every other tag
change, so it is one write rather than one per thread.
Disabled until the query reports its total. Threads arrive in batches,
so before then the model holds only what has landed, and an action
saying "all" must not silently skip the rest. A greyed control says
"not yet" without needing a dialog or a stall the user cannot see.
The state is also set at registration, since QAction starts enabled
and a window that has not run a query has nothing to act on.
Two things came out differently from the plan, both forced by existing
code. It carries a default binding, because everyActionHasAShortcut
requires every registered action to have one: an unbound action is
unreachable from the keyboard, and that invariant is deliberate, so the
action was given Ctrl+Shift+U rather than the invariant relaxed. And
only the threads that are actually unread are sent, because sending the
rest would inflate the pending-edit count with writes that change
nothing, and the quit prompt reads that count. A view with nothing
unread does nothing, pushes no command and says so: an undo entry that
restores nothing is worse than none, since it absorbs a Ctrl+Z meant
for the previous action.
undoDepthForTesting() is new and exists for a reason worth recording:
undo->isEnabled() cannot answer "was a command pushed", because the
undo QAction is always enabled and tests canUndo() when triggered. The
first version of the no-op test asserted on it and passed against a
mutant with the unread filter removed.
Closes item 43.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
"Syncing..." was set once and never updated, so a run that takes over a
minute reported nothing about what it was doing.
The original diagnosis in the backlog was half wrong, and two further
wrong ones were made and discarded before the real cause: plain
"mbsync -a" prints NOTHING until it exits, then a single summary line.
Measured on a real run, one line at 11:11:08 then 73 within the second
11:11:33, at the end of a 46-second run. So there was no stream to read
for the part of a sync that takes time. It is not buffering, so stdbuf
changes nothing, and the account name is not unavailable either, which
was the second wrong conclusion.
mbsync -V is what changes both: it announces each channel as it reaches
it, which is at once the progress and the account name originally
asked for. The shipped script now passes it.
SyncPhaseTracker derives a short status from the output as it streams:
the channel being synced, the summary counts when mbsync ends, then the
notmuch reindex. It lives beside MailSync rather than in the window so
the matching rules are one testable thing, and it holds no widget.
Matching is loose and case-insensitive, since the wording varies by
version, and nothing in it decides success or failure: the exit status
remains the only authority on that.
Lines are reassembled in MainWindow before being fed, because
QProcess::readAll() splits wherever it happens to and a half-line would
match nothing. Every status is sanitised and truncated: the channel
name comes from a config file this app does not own, and a long one
must not stretch the status bar.
Two defects in existing code, fixed with it. setSyncBusy(true) ran
after start(), so a fast run's output arrived before the per-run reset
and wiped its own phase. And a first draft deferred phases while a
transient message showed, which let a "Background sync completed"
message armed before the sync began suppress the whole run: a running
sync's state outranks an expiring event message.
Verified by replaying real captured mbsync -V output through the
tracker, not only against fixtures. The MainWindow test paces its
script with sleeps, since a script that prints everything at once
arrives in one readyRead and makes every intermediate phase
unobservable.
Closes item 42.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Sync had two controls that behaved differently. The QPushButton beside the
query bar cleared the log, opened the pane and disabled itself; the QAction
behind the toolbar, menu and shortcut called start() and did nothing else,
discarding its return value so a rejected start was silent. Worse, item 29's
"disable Sync during a background sync" set the button only, so the toolbar
entry stayed clickable through a cron sync and could only produce the
script's EX_TEMPFAIL skip.
startSync() is now the single handler behind every route in, and the enabled
state lives on the QAction, which reaches the toolbar, the menu and the
shortcut at once. It also reports when no sync command is configured rather
than doing nothing.
The QPushButton is gone. It read as a Search button given it sat beside a
text field, which is the user's own observation and the reason the toolbar
one survives instead. Its unavailable-command tooltip moved to the action,
since that is the only thing that says why the control is dead.
Removing it left the query field running flush to the window edge, so the
saved-query buttons move from their own row onto the query row. The bar is
now framed by the account dropdown on the left and the saved queries on the
right, the empty row is gone, and the thread list gains the space. The field
also gains setClearButtonEnabled, which is Qt's own themed clear icon rather
than a hand-rolled button. A "Search" button was considered and rejected:
Return already runs the query.
No overflow handling for [queries], which is unbounded. Three entries fit;
item 23 already specifies buttons-plus-menu and is where that belongs.
CLAUDE.md's architecture diagram named four widget classes that have never
existed, QueryBar, SavedQueryBar, HeaderWidget and AttachmentBar. The query
row and the message header are built inline. Corrected, and the components
that do exist but were missing from it added.
Tests: the new action test was verified red first and load-bearing by
mutation. The old button test is deleted rather than repointed, being an
exact duplicate of it, and the unobservable-lock test now drives the action.
The clear button and the row layout were confirmed by hand; no test clicks
the icon, which is a mouse path.
Backlog: 45 done and reclassified as a defect rather than a cosmetic
redundancy, 47 added for the bar.
|
|
Both are the same class of defect: a test that reads real machine state and
so passes or fails on circumstance rather than on the code under test.
Item 38. Every MainWindow a test builds constructed its SyncMonitor on the
live /proc/locks, so a window observed the machine's actual sync state and
the sync-button assertion failed whenever the user's cron sync happened to
be running. Cron fires every ten minutes and a run lasts ~35s, which is
roughly 6% of runs, and it read as flakiness. SyncMonitor already took an
injectable locks path for exactly this; MainWindow did not expose it. It
does now, as a test seam rather than a config key: /proc/locks is not
something a user would set, and a wrong value silently disables background
sync detection instead of failing loudly.
The monitor is still constructed and started, per the item's own constraint.
Only the table it reads is redirected, to an empty file in the test's own
temporary directory.
Item 46. uiStateSurvivesARestart asserted a 940px width, and the offscreen
platform reports an 800x800 screen. restoreGeometry() clamps to the
available area, so the width came back as 798 while the 620 height, which
fits, restored untouched. That asymmetry was the tell that persistence was
fine and the test was wrong. The size is now 640x560 and carries no meaning
beyond differing from the default.
Verified by reproducing the original conditions rather than by waiting for
them: the suite run under flock -n /tmp/mbsync.lock fails item 38's
assertion with the seam bypassed and passes with it in place, and item 46
now passes under offscreen where it failed. One dud mutation is recorded in
the backlog, writing an unparseable line into the injected lock table does
not fail the test, because lockHeldIn() correctly finds no lock in it.
Suite: 15/15 offscreen with the lock held, and green on Wayland except the
pre-existing querycompleter screenshot flake, which fails to grab under
Wayland and passes offscreen.
|
|
A tag edit sent while another process holds notmuch's write lock does not
fail: the read-write open blocks and then succeeds. Measured against
Slackware's notmuch, 9.158s against a 12s hold, status SUCCESS. Since the
worker is a single thread, that blocked open holds up every read queued
behind it, so the message pane freezes on whichever thread was selected
first and replays the queue when the lock releases.
The window now defers instead. While SyncMonitor reports a sync running, a
tag change is held rather than sent, and flushed when the sync ends. The
optimistic update stands in the meantime, so the row keeps its tag and the
edit still counts toward the unsynced indicator, which is what the quit
prompt reads.
The original diagnosis was that the open fails and the edit is discarded,
and a retry was built on it. That was wrong: the error branch in
notmuchworker.cpp is unreachable through lock contention. The premise was
taken from a plausible-looking error path without provoking the condition,
and measurement disproved it. The backlog entry records this rather than
quietly correcting it.
Verified by hand against a real blocking open, which the tests cannot reach:
they drive the deferral through the meta-object and never take a lock. Both
locks held for 100s with a tag edit made during the hold. Row kept the tag,
status did not expire, indicator rose, window stayed responsive, held edit
sent itself on release.
The 2s SyncMonitor polling window is knowingly left open: a sync starting
between polls is invisible for up to 2s and an edit there still blocks.
SyncMonitor::lockHeldIn() would close it at the cost of one file read per
tag action, and is recorded as the option to revisit.
Also fixes revertPendingTagChange() clearing the entire undo stack after any
rejected write, found while working on this.
Backlog: item 37 done, and item 46 added for a test that fails only under
the offscreen platform, where an 800x800 screen clamps a restored 940px
window. Pre-existing and unrelated; the suite is green otherwise.
|
|
Item 28, reported by the user: open a thread, let the automatic
mark-read remove `unread`, then press Ctrl+U to put it back. The
indicator read "2 unsynced change(s)" with the mail store exactly where
it started.
The counter incremented per confirmed write and never decremented, so
any add-then-remove of the same tag inflated it. Mark-read is simply the
path that fires without being asked, which is why it surfaced there.
The user's call was net state: an edit and its inverse are zero
outstanding changes, because what the indicator answers is whether
quitting now would strand work. A QHash keyed "<messageId>\n<tag>"
replaces the int, and a pair that reverts is erased rather than stored
with the new direction, so the map cannot grow without bound across a
long session of tagging and untagging.
Keyed per (message, tag) rather than per message: removing `unread` and
adding `flagged` on one message are independent changes and must not
cancel each other. A change carrying no message ids cannot be netted
against anything and is counted separately, since dropping it would
understate the indicator, which is the direction that costs work.
Both properties item 18 established still hold: a successful sync clears
everything, a failed one clears nothing.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Items 16 and 33.
Delete now removes the `deleted` tag when every selected thread already
carries it, so pressing it twice puts a thread back. One direction for
the whole selection, never per row: toggling each independently would
leave a single keystroke with the selection in two states, which is
worse than either outcome.
Status messages are classified rather than blanket-timed, which is the
substance of item 33. Events expire after six seconds and fall back to
the last query's thread count: "Sync complete", "Nothing to undo", the
skip notice, the per-action "Archive: 3 threads". State does not expire:
"Searching...", "Syncing...", the selection count, and "Sync failed
(exit N)", because an error must not vanish before it is read.
A test caught a mistake in that routing. Making the per-action message
transient armed the timer during select-all, since tagSelected() runs on
a selection onSelectionChanged() had just described, and the count would
then be replaced while it was still true. Writing the count now cancels
any transient still counting down.
QStatusBar::showMessage() would give the same behaviour but the label is
added with addWidget() beside permanent widgets, so adopting it means
reworking that arrangement. One timer beside the label is the smaller
change.
Both fixes verified by reverting them and watching the tests fail.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Items 29 and 32.
29 was a constraint item 27 specified and that shipped unbuilt: while a
cron sync held the lock the Sync button stayed clickable, and pressing it
could only produce the EX_TEMPFAIL skip. The progress bar and the button
are now written by one updateSyncControls() taking both sync sources,
which the item asked for by name: two independent assignments, one per
path, means whichever finishes second wins, so a background sync ending
would re-enable the button in the middle of a local run.
Unknown re-enables the button, deliberately. It means /proc/locks could
not be read and nothing was observed, so leaving the button disabled
would strand it permanently wherever the lock cannot be seen.
32 adds a clear_pane action on Esc. It clears m_currentThreadId with the
pane, not merely alongside it, or a threadLoaded still in flight would
paint the thread straight back; and it cancels any pending mark-read,
since a thread blanked from view must not be marked read two seconds
later. The selection, the query and the undo stack are untouched.
The one real risk in 32 was Escape being stolen from the query
completer, the way Return was once lost to a window shortcut. Probed
rather than reasoned about: a popup consumes the key before a
window-level shortcut sees it, so the completer still dismisses.
Every test here was verified by reverting the code it covers.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Reported from hand testing: a manual sync ended with "Sync finished
elsewhere" stamped over its own result.
Ownership of a lock period was being decided when the lock was RELEASED,
by asking MailSync::isRunning(). That question cannot be answered then:
the process exits, so isRunning() goes false, and only afterwards does
the next poll observe the lock gone. The guard therefore suppressed the
message while the sync ran and let it through at the end, up to two
seconds after onSyncFinished() had already said what happened.
Ownership is now latched when the lock APPEARS, which is the moment
isRunning() can still answer, and the matching release is swallowed.
onSyncFinished() hands the latch back when it sees exit 75, because a
skip means the lock was never ours: if a manual run and the cron run
start inside one poll interval, the lock would otherwise be latched as
local and that other run's completion swallowed with it.
Also renames the messages to "Background sync running/completed" per the
user: "finished elsewhere" reads as though the application does not know
what is syncing the Maildir, when in fact it is the same script.
The tests added here cover the external path and the Unknown state. They
do NOT reproduce the reported bug, and were checked against a reverted
fix to confirm that: staging it needs isRunning() true at the Running
transition and false at the Idle one, which cannot be arranged in
test_mainwindow without a configured sync command and a live child
process. That was tried and abandoned, it left a process running for the
length of the suite and popped a dialog. The ordering and the fix were
instead verified against a standalone model of both code paths.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
The user's cron runs mailsync.sh every ten minutes, so mail arrives and
tags change while the window sits idle, and nothing here noticed. The
script already holds an flock for the whole run, so that lock is the
signal: no status file is needed, and a kernel lock cannot go stale
because it dies with the process holding it.
The observation method is the part that matters, and two of the three
plausible ones are wrong. Both were probed on Linux 6.18 before any of
this was written:
- flock -n acquires in order to test, so polling every two seconds
would open a window every two seconds in which a starting
mailsync.sh is refused the lock and exits 75. It would cause the
very skips the script reports.
- fcntl(F_OFD_GETLK) never acquires and looks ideal, but reports
UNLOCKED against a lock held by flock(2): separate lock namespaces
in the kernel, which cannot see each other. A silent false negative.
- /proc/locks is a pure read. It observes flock(2) correctly, and
since it takes no lock at all it can never contend with the Xapian
write lock notmuch new holds during the same run. Confirmed: 200
reads left the lock table unchanged and this process holding
nothing.
SyncMonitor keeps the parsing separate from the polling so the parsing
is testable, and it reports Unknown rather than Idle where /proc/locks
cannot be read: "no sync is running" is the claim that would let the
window quit, so it must never be guessed. Verified against a real flock
end to end, not only against synthetic content.
It reports rather than refreshes. runCurrentQuery() clears the undo
stack, the selection and the message pane, which is right for a query
the user typed and hostile for one a cron timer fired: it would discard
undo history and close the thread being read up to six times an hour,
with no action from the user.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Multi-select already worked by Ctrl+click and Shift+click, but nothing in
the UI said so and every tag action was keyboard-only, so the Ctrl+T tag
dialog could not be reached with a mouse at all.
Adds a select_all action on Ctrl+A, registered like every other action so
it reaches the Edit menu, the shortcut reference and [keys]; a right-click
menu on the thread list built from the same QActions rather than parallel
copies; a selection count in the status bar, which is the part that
actually teaches the feature by acknowledging a selection while it is
being built; and a note in the shortcut dialog for the mouse gestures,
which are view behaviour and cannot appear in the generated table.
A selection gesture must not open mail or mutate it. Selecting several
rows now blanks the message pane and cancels any pending mark-read,
rather than rendering each row swept through and queueing it to be marked
read.
Two Qt behaviours shaped this, both established by probe rather than from
memory:
- selectAll() emits no currentRowChanged at all and leaves the current
index invalid.
- currentRowChanged is emitted BEFORE the selection model is updated.
The second one caused two distinct faults. Collapsing a multi-row
selection back to one row reported the old count, so the guard swallowed
the load and the pane stayed blank; that case is handled in
onSelectionChanged, which sees the true count. And a Ctrl+click taking
the selection from one row to two also reported one, so the thread was
loaded, blanked, and then painted back when the queued reply returned
from the worker. By the third row the id was already cleared and the
reply was discarded, which is why the fault presented as an off-by-one in
the threshold rather than as a race.
Tests cover the synchronous half. The late-reply guard has no test:
MainWindow in tests has no worker, so threadLoaded never fires and the
repaint cannot be reproduced in process. Verified by hand instead.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Four changes to the sync UI, three of them from using it.
The log pane could not be dismissed. It is hidden at construction and
shown on failure, and nothing ever hid it again, so a single failed sync
left it on screen until the application restarted. It now sits in a
container with its own Close button, and is 200px rather than 120,
because mbsync's output is wide and repetitive and the shorter pane
showed too little of it to read. It still appears only on failure, which
the user confirmed is what they want.
A sync gives no feedback while it runs. The status bar now carries an
indeterminate progress bar for the duration, and the Sync button is
disabled rather than left looking live. The bar is indeterminate on
purpose: mbsync reports no percentage and the script's output is
unstructured, so a bar filling left to right would be inventing a
fraction nobody knows. The log is also cleared at the start of each run,
since leaving the previous run's lines in place makes a stale failure
look like the current one.
The lock skip was reported as a failure. mailsync.sh exits when another
run holds the lock, and that was exit 1, which qtmaildir reads as "sync
failed": it showed the log pane and, on the exit path, told the user
their changes were still unsynced. With a cron timer every ten minutes,
a click landing inside a run is routine and none of that is true. The
script now exits 75 (EX_TEMPFAIL) and the window reports it as its own
case, saying a sync is already running. On the exit path it stays open
and says plainly that the other run is most likely carrying the changes
over but that this window cannot see it finish, rather than guessing
either way.
That last hedge is what item 27 records: the application cannot see a
sync it did not start. The user chose continuous polling of the lock file
over the narrower "only while quitting" version, and the entry notes that
the lock is already the signal, so no status file is needed, and that a
kernel lock cannot go stale where a written file can.
Verified against stub binaries: a second run while the lock is held exits
75 and says SKIPPED, while the run holding it completes at 0.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Five hardcoded tags were the only ones reachable from the UI: archive,
delete, spam, flag and toggle_unread. For an application whose purpose is
organising mail by tag, applying any other one meant leaving for a
terminal. Item 26 of the usability backlog, raised by the user asking how
to add a tag and finding they could not.
One dialog rather than separate add and remove actions, at the user's
choice: filing something under a new tag while dropping inbox is one
thought, not two. Type tags to add or remove, comma separated, or clear a
checkbox to drop a tag already on the selection without retyping its
name.
Both fields complete against the tag list MainWindow already holds for
the query completer. Completion is a guard against typing shoppping
beside shopping, never a whitelist: inventing a tag is the entire point,
so any valid name goes through whether or not it exists yet.
Tri-state checkboxes carry the multi-thread case, and are where the risk
is. A tag on some selected threads shows partially checked, and leaving
it alone changes nothing; the opposite reading would silently tag threads
the user never looked at. A tag already on every thread and left checked
is likewise not a change and is not sent as one.
Tag names are validated before anything is applied, through a free
function so the rules are testable on their own. Empty, a leading dash
(notmuch's CLI reads it as removal, making such a tag a trap), whitespace
and control characters are refused by name and reason. Nothing is applied
until the whole set passes, since the user cannot tell which half of a
partial change landed.
TagDialog is pure UI: handed the vocabulary and the current state,
returning two lists, contacting no worker. That is what lets its fifteen
tests run without a notmuch database. Integration is a single call to the
existing tagSelected(), so undo, the optimistic model update, the
combined multi-row query and the completer refresh for a brand-new tag
all come for free.
One test assumption was wrong and the code was right: a case asserted
that QStringLiteral("null\0byte") truncates at the null and reads as
empty. It does not, so the null is caught as a control character. The
test was corrected rather than the validator.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Tagging changes the notmuch index at once, but the mail store only hears
about it on the next sync, and nothing said so. Quitting with tagging
outstanding was silent. Items 18 and 19 of the usability backlog, built
together because the second needs the first's counter.
The counter cannot be QUndoStack::isClean(), which is the obvious
candidate and the wrong one: the undo stack is cleared on every query,
since its entries refer to rows the new result set discards. Tag a
thread, run any query, and the stack is empty while the change is still
unsynced. m_pendingEdits is its own count, incremented where a write is
CONFIRMED rather than where one is sent, so an optimistic update the
worker later rejects cannot leave the indicator claiming an edit that
never landed. Only a successful sync resets it: clearing on failure would
assert the changes had reached the mail store when the sync is exactly
what failed to put them there.
It is shown in the status bar, hidden entirely at zero, and described as
a lower bound rather than a guarantee, since an external notmuch run can
carry changes over without this application noticing.
On exit, sync_on_exit in [general] takes ask, always or never. Three
values rather than a bool because "prompt me", "just do it" and "do
nothing" are three behaviours and true/false expresses two; an unknown
value warns by name, since a typo there silently changes what happens to
unsynced work. The prompt offers three buttons for the same reason: a
user who hit Quit by mistake needs a way back that is not "sync". A sync
started at exit holds the window open until it finishes rather than being
killed mid-run, and a sync that FAILS does not quit, because quitting
there would discard the user's choice silently. With no sync command
configured the prompt degrades to a plain warning instead of offering a
sync that cannot run.
This is not a destructive-action confirmation of the kind CLAUDE.md
forbids. Those cover tag mutations, which keep undo instead of a dialog.
This asks about losing work at the one point where undo cannot help.
The tagsApplied lambda became a named slot, which is better structure and
also what lets a test drive it: the worker is deliberately parentless
because it moves to its own thread, so reaching it with findChild to emit
the real signal cannot work, and contorting the test to try was the wrong
instinct. Testing a modal needed its own care. A test that sends a close
event hangs forever if an unexpected dialog opens, because the modal
spins its own event loop; CloseProbe polls for activeModalWidget, closes
it and records that one appeared, turning "a dialog opened" into an
assertion rather than a hang.
Also removes a stray qDebug left in the open_thread action by the earlier
Enter-key investigation, which had reached two commits.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Opening a thread left it tagged unread, so the unread count never
matched what had actually been read and the app was quietly wrong every
day it was used. Item 6 of the usability backlog.
A single-shot timer, armed when a thread is selected and restarted rather
than stacked, so arrowing down a list marks only the thread still
selected when it fires and not every one passed through. Configurable
through mark_read_delay_ms in [general], defaulting to 2000: zero marks
read at once, and any negative value disables the behaviour, which is why
the value is neither clamped nor warned about at either end.
The automatic change deliberately does NOT go on the undo stack. It
routes through sendThreadTagChange() rather than tagSelected(), because
undoing an action the user never took is worse than leaving a thread
read, and toggle_unread already gives them a direct way back. It still
funnels through the single applyTags path; what differs is only whether
the inverse is pushed, which is a window-level decision above the worker.
An explicit toggle_unread cancels any pending timer, or marking a thread
unread by hand would be reversed a moment later and the key would look
broken.
Two guards beyond the plan, both from asking what happens when a timer
outlives the thread it was armed for. Arming is skipped for a thread that
is not unread, so no write is scheduled that would change nothing, and
the handler re-checks that its thread is still selected and still unread
before writing, so a stale timer does nothing rather than tagging the
wrong thread.
The plan expected the rapid-arrow case to need a database and a manual
check. It needs neither: ThreadListModel takes threads through
appendBatch(), so the case is unit-tested. All three tests were confirmed
to fail against deliberately broken versions, one arming for read threads
and one creating a timer per selection instead of restarting one.
Item 7 is closed in the same pass. The user verified against real mail
that HTML messages already open as HTML, which is what the item asked
for, so it is recorded as done with no code changed. The prefer_html key
it floated was not added: nobody has asked to default to plain text, and
Ctrl+H already switches a thread by hand.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Pressing Return in the query bar moved focus to the thread list and left
the query unrun, so a query typed at the keyboard could not be executed
at all.
Return is bound to open_thread as a Qt::WindowShortcut, and a shortcut is
dispatched before the focused widget ever sees the key. Qt withholds a
plain-LETTER shortcut from an editable widget, which is why every other
binding in the map was safe here, but Return is not a letter and gets no
such protection: the action fired from inside the bar, its handler called
setFocus() on the thread list, and QLineEdit::returnPressed was never
emitted.
Accept the ShortcutOverride for Return and Enter on the query bar, which
tells Qt the focused widget wants the key as ordinary input and stops the
shortcut being dispatched. Narrow by design, one widget and one key, so
open_thread keeps working everywhere else in the window.
Three earlier hypotheses were tested and disproven before this one, and
two synthetic probes wrongly reported the binding as harmless: real input
sends ShortcutOverride first and only dispatches the shortcut if nothing
claims it, while QTest::keyClick skips that round trip entirely. The new
test drives the override exchange rather than the keystroke and fails
against the old code. The comment claiming no filter was needed said the
letter rule covered this case; it did not, and it now says so.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Adds the complete_query action, whose registration the assertion in
registerActions() has been demanding since keymap gained the name: a
known action nothing implements would have been a silently dead binding.
Also lands the QueryCompleter half of the manual trigger, which the
keymap commit left behind: triggerCompletion() and the focus-in filter
gated on completion_on_focus.
Tags refresh at startup, after a sync, and after a mutation introduces a
tag not already known, since that is the tag most likely to be typed
again. onAllTagsReady deliberately drops the generation the signal
carries: a tag list is not an ordered query result, so a later one is
always at least as good as an earlier one and discarding on staleness
could only throw away a good list.
|
|
Resizing the window, the splitter or a thread-list column was undone by
the next launch. State now round-trips through a separate settings file
at ~/.local/state/qtmaildir/uistate.conf, written on close and read at
startup.
The state file is deliberately not the user's config: a base64 geometry
blob does not belong in a hand-edited file, and rewriting that file on
exit would drop its comments and key order, which QSettings does not
preserve.
Two details that are easy to get wrong:
QStandardPaths::StateLocation appends both the organization and the
application name, and both are "qtmaildir" here, so it resolves to
~/.local/state/qtmaildir/qtmaildir. The path is built from
GenericStateLocation instead, matching Config::defaultPath().
Restore runs after buildMenus() rather than at the end of buildUi():
QMainWindow::restoreState() matches toolbars by object name and silently
drops the position of one that does not exist yet.
Every restore is conditional on a non-empty blob, so a missing or
rejected state file leaves the built-in defaults instead of producing a
zero-size window.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Two optional keys in an [account.<key>] stanza. color fills the chip,
label sets its text.
Both belong to the account rather than to [tagcolors] because an account
tag is a different taxonomy: which mailbox a thread arrived in, not what
state it is in.
label is display only. "account-provider-work" is a lot of
row for one bit of information, but the notmuch tag is never renamed, so
existing queries and external tagging are unaffected. Unset falls back to
the account key, and an empty label is ignored rather than rendering a
blank chip.
|
|
Actions were a QHash of std::function dispatched by an event filter,
which nothing could put in a menu. They are QActions now, bound from
KeyMap so a [keys] override reaches the menus as well as the keyboard.
Menu bar covers every action; the toolbar carries only Sync, Archive,
Delete and Undo. Help > Keyboard shortcuts is generated from the actions,
so it shows what the keys really do rather than a copy that drifts.
spam and load_remote gained defaults, having been unreachable without a
hand-written binding.
The event filter is gone. Probing showed QAction shortcuts are dispatched
before the focused widget sees the key, so they beat QAbstractItemView's
type-to-search without one, and Qt already suppresses plain-letter
shortcuts while an editable widget has focus. Dropping the filter's
blanket guard also lets Ctrl+Q work while the query bar has focus.
registeredActionNames() is derived from the actions rather than
hand-maintained, so the two drift tests it needed are replaced by checks
that a configured binding reaches its action.
No confirmation dialogs: tag mutations still answer to undo.
|
|
Confirmed with the maintainer as v2-only rather than v2-or-later. LICENSE is
the official text from gnu.org. Every file under src/ and tests/ carries the
matching notice.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
|
Wires the worker thread, thread list, message pane, sync process and undo
stack together, and replaces the placeholder main() with real startup:
custom URL schemes registered before QApplication, a libnotmuch ABI check,
and config loading.
Four fixes against the drafted version:
- onWorkerError() only set a status label. Its own comment elsewhere
claimed it reverted the optimistic update, and the spec requires that;
it did not, so a rejected write left the list showing a tag the database
never received. The pending change is now recorded and rolled back, and
a confirmed tagsApplied clears it so a later unrelated error cannot undo
a write that succeeded.
- runCurrentQuery() cleared the model but left the undo stack pointing at
rows that no longer exist. Undoing after a new query would have written
to the database while the visible list stayed put. The stack is cleared
with the model.
- m_currentMessages was assigned on every thread load and never read.
Removed.
- buildUi() connected sync output to m_syncLog and errors to m_statusLabel
before either existed. Both are constructed before the wiring now.
cidPrefix generation lives here, this being its only producer in the
application, and is pinned by tests: it must never contain '!' and must be
distinct per message, which are the invariants the cid: namespacing rests
on. A second test holds registeredActionNames() against
KeyMap::knownActions(), since those two hand-maintained lists drifting
either way silently breaks a user's key binding. Mutation-verified that
dropping an action fails the test by name.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
|
