In the recent changes for search order handling the default-value of
notmuch-search-oldest-first was used. However, default-value needs a
symbol so the symbol-name needs to be quoted.
This missing quote was causing strange sort-orders in some cases.
The only user-visible effect of this should be that "G" now works in
show mode (previously it was unbound for no apparent reason).
This shared keymap gives us one place to put global commands, which
both forces us to think about what commands should be global, and
ensures their bindings can't diverge (like the missing "G" in show).
This converts notmuch-help to use map-keymap for all keymap traversal.
This generally cleans up and simplifies construction of keymap
documentation, and also makes notmuch-help support anything that can
be in a keymap, including more esoteric stuff like multiple
inheritance.
This unifies the various refresh and poll-and-refresh functions we
have for different modes. Now all modes bind "=" and "G" (except
show, which doesn't bind "G" for some reason) to
`notmuch-refresh-this-buffer' and
`notmuch-poll-and-refresh-this-buffer', respectively.
Since notmuch-hello doesn't need this any more, we can remove this
hack. This also eliminates `notmuch-search-quit', so now all modes
bind "q" to `notmuch-kill-this-buffer'.
Since there is now no difference between notmuch-hello-search and
notmuch-search when called interactively, bind "s" to notmuch-search
in notmuch-hello-mode-map. Now all modes bind "s" this way.
Previously, we refreshed hello when the user quit a search that was
started from hello. This is fine assuming purely stack-oriented
buffer use, but is quite fragile and requires hacks to search.
This replaces that logic with a new approach that refreshes hello
whenever the user switches to the hello buffer, regardless of how this
happens.
Previously, if `notmuch-search' was called interactively (bound to "s"
in search and show, but not hello), it would always use newest-first.
However, `notmuch-hello-search' (bound to "s" in hello) and
`notmuch-hello-widget-search` would call it with the user-configured
sort order. This inconsistency seems unintentional, so change
`notmuch-search' to use the user-configured sort order when called
interactively.
When text/html parts include images as multipart/related and the
text/plain alternative is used these images can be completely hidden
with no easy way to access them or even find out that they are there.
Make notmuch-show-insert-part-multipart/related add buttons for all
parts, the first one visible the rest hidden.
The lazy part handler had a bug that it allowed the button to be
toggled to be specified. During toggling it needs to save and restore
the text-properties for the button but it actually saved the text
properties at point rather than from the button.
In almost all cases this didn't matter as as point had the same text
properties as the button. However, it is a bug and did cause incorrect
behaviour in some cases: see id:87txhz14z6.fsf@qmul.ac.uk for details.
When composing a reply, notmuch-mua-reply attempts to cite the
the original message by inserting it before the user signature, if
one is present. The existing method used to search the signature
separator backward from the end of the buffer and then move one
line up. In case of variable `message-signature-insert-empty-line'
being nil this caused point to go to the beginning of
'--text follows this line--'
separator line, and citation was inserted there.
This change checks the value of `message-signature-insert-empty-line'
and doesn't move point if that is nil. Additional narrowing to
the body region ensures that point never goes to the separator line
(or beyond).
`message-signature-setup-hook' or `message-setup-hook' may already have
added some other content to the message body, therefore using simply
(message-goto-body) to move point to the beginning of body might lead
to unexpected results.
Original patch from "Geoffrey H. Ferrari", continued with iterations
from Jani and Mark.
In reply, insert quotable parts using notmuch-show-insert-bodypart
instead of calling notmuch-mm-display-part-inline directly to render
the quoted parts as they are rendered in show view.
We use a temp buffer to not leak text properties from the show
renderer into the reply. This way we also don't need to worry about
narrowing or point placement. Credits to Mark Walters
<markwalters1009@gmail.com> and Austin Clements <amdragon@MIT.EDU> for
getting this part straight.
The notable change is that replies to text/calendar parts quote the
pretty printed output of icalendar-import-buffer rather than the ugly
raw vcalendar.
notmuch-show.el and notmuch.el had 3 duplicate, identical functions:
notmuch-foreach-mime-part, notmuch-count-attachments and
notmuch-save-attachments. Now these functions in notmuch-show.el
are replaced with declare-functions pointing to "notmuch"(.el).
Notmuch puts attachments in as declared content-type except when the
content-type is application/octet-stream it tries to guess the type
from the filename/extension. This means that viewing a pdf (for
example) which is sent as application/octet-strem invokes the pdf
viewer rather than just offering to save the part.
Recent changes to the attachment handling (commit 1546387d) changed
(broke) this. This patch stores the calculated mime-type with the part
and changes the attachment part handlers can use it instead.
Recently notmuch-hello was converted to use batch count. However, it
seems that several people run different versions of notmuch-emacs and
notmuch-cli so this batch makes emacs fail with an error message if
--batch is not available in the CLI.
Amended by: db
Notmuch cli provides all structured data previously provided
in json format now in s-expression format, rendering all current
json functionality obsolete.
This switches `notmuch-mua-reply' and `notmuch-query-get-threads' to
the S-exp format. These were the last two uses of the JSON format in
the Emacs frontend.
This is just like `notmuch-call-notmuch-json', but parses S-expression
output. Note that, also like `notmuch-call-notmuch-json', this
doesn't consider trailing data to be an error, which may or may not be
what we want in the long run.
This removes the v command, since we now have much nicer part commands,
and deprecates the underlying notmuch-show-view-all-mime-parts. This
also means that people who try using the old unprefixed 'v' command on
a part button will no longer be greeted by ALL of their parts popping
up.
Previously the query string for piping a message to a command was
"Pipe message to command: " regardless of whether the function was
called with a prefix argument (which pipes all open messages to the
command). This patch modifies the `interactive' command to reflect
this.
This adds the actual code to do the lazy insertion of hidden parts.
We use a memory inefficient but simple method: when we come to insert
the part if it is hidden we just store all of the arguments to the
part insertion function as a button property. This means when we want
to show the part we can just resume where we left off.
One thing is that we can't tell if a lazy part will produce text until
we try to render it so when unhiding a part we check to see if it
rendered; if not we invoke the default part handler (e.g. an external
viewer).
Also, we would like to insert the lazy part at the start of the line
after the part button. But if this line has some text properties
(e.g. the colours for a following message header) then the lazy part
gets these properties. Thus we start at the end of the part button
line, insert a newline, insert the lazy part, and then delete the
extra newline at the end of the part.
Previously, whether a part was hidden or shown was recorded in the
invisibility/visibility of the part overlay. Since we are going to
have lazily rendered parts with no overlay store the hidden/shown
state in the part button itself.
Additionally, in preparation for the invisible part handling move the
actual hiding of the hidden parts to insert-bodypart from
create-part-overlays.
Finally, we will need to know whether a part-insertion has done
anything (it won't if the invisible part cannot be displayed by emacs)
so we slightly rejig the code order in
notmuch-show-toggle-part-invisibility to make it easier for the
function to set an appropriate return value.
Previously each of the part insertion handlers inserted the part
button themselves. Move this up into
notmuch-show-insert-bodypart. Since a small number of the handlers
modify the button (the encryption/signature ones) we need to pass the
header button as an argument into the individual part insertion
handlers. However, the declared-type argument was only used for the
text for the part buttons we can now omit it.
The patch is large but mostly simple. The only things of note are that
we let the text/plain handler applies notmuch-wash to the whole part
including the part button. In particular, notmuch-wash removes leading
blank lines from a text/plain part, but since the button is counted as
part of the part this does not happen with text/plain buttons that
have a button. This is probably a bug in notmuch-wash but changing it
does make several tests fail (that rely on this blank line) so, for
the moment, keep the old behaviour.
Earlier patches have moved the handling of wash fake inline patch
parts to insert-bodypart so we can drop the function
notmuch-show-insert-part-inline-patch-fake-part
Occasionally, when the user killed the search buffer when the CLI
process was still running, Emacs would run the
notmuch-start-notmuch-sentinel sentinel twice. The first call would
process and delete the error output file and the second would fail
with an "Opening input file: no such file or directory, ..." error
when attempting to access the error file.
Emacs isn't supposed to run the sentinel twice. The reason it does is
rather subtle (and probably a bug in Emacs):
1) When the user kills the search buffer, Emacs invokes
kill_buffer_processes, which sends a SIGHUP to notmuch, but doesn't do
anything else. Meanwhile, suppose the notmuch search process has
printed some more output, but Emacs hasn't consumed it yet (this is
critical and is why this error only happens sometimes).
2) Emacs gets a SIGCHLD from the dying notmuch process, which invokes
handle_child_signal, which sets the new process status, but can't do
anything else because it's a signal handler.
3) Emacs returns to its idle loop, which calls status_notify, which
sees that the notmuch process has a new status. This is where things
get interesting.
3.1) Emacs guarantees that it will run process filters on any
unconsumed output before running the process sentinel, so
status_notify calls read_process_output, which consumes the final
output and calls notmuch-search-process-filter.
3.1.1) notmuch-search-process-filter checks if the search buffer is
still alive and, since it's not, it calls delete-process.
3.1.1.1) delete-process correctly sees that the process is already
dead and doesn't try to send another signal, *but* it still modifies
the status to "killed". To deal with the new status, it calls
status_notify. Dun dun dun. We've seen this function before.
3.1.1.1.1) The *recursive* status_notify invocation sees that the
process has a new status and doesn't have any more output to consume,
so it invokes our sentinel and returns.
3.2) The outer status_notify call (which we're still in) is now done
flushing pending process output, so it *also* invokes our sentinel.
This patch addresses this problem at step 3.1.1, where the filter
calls delete-process, since this is a strange and redundant thing to
do anyway.
Previously, when the user killed the search buffer before the CLI
search process had completed, we would report the signal sent by Emacs
to kill the CLI to the user as an error. Fix this by only reporting
error exits if the process buffer is still live. We still report
stderr output regardless in case stderr output was relevant to why the
user killed the search buffer (such as a wrapper script being stuck).
This commit adds an extra button at the end of the search entries that
allows deleting that individual search from the history. A short
confirmation («y» or «n») is made before taking action.
The button to clear the recent searches in notmuch-hello is easy to
press accidentally while moving around the, clearing potentially
useful searches with no way of recovering them.
Previously, we simply called pushnew to add :notmuch-part to the
front-sticky and rear-nonsticky text property lists. This works if
these are nil or lists, but they can also have the value t, meaning
that all properties are front-sticky/rear-nonsticky. In this case,
pushnew will signal an error because t is not a list. We never set
these properties to t ourselves, but since we apply these property
changes over arbitrary renderer output, we have to deal with this
possibility.
Several function docstrings refer to behaviour in docstrings that is
really controlled by notmuch-archive-tags. Add cross references, and
replace hardcoding.
These functions refer to default values of variables, but it seems
less confusing and less likely to get out of date to just allow the
user to follow the help cross-reference links.
I found several places where a setq is immediately followed by a let
or a let*. This seems to be the pessimal combination, with the
implicit scope of the setq combined with the extra indentation of the let.
I combined these cases into a single let* which I think is easier to read.
I can't see any benefit to the funcall, and it looks like the result
of cut-and-paste from some code that actually used a variable for the
function to call.
In addition to being the Right Thing to do, this noticeably improves
the time taken to display the first page of search results, since it's
roughly an order of magnitude faster than the JSON parser.
Interestingly, it does *not* significantly improve the time to
completely fill a large search buffer because for large search
buffers, the cost of creating author invisibility overlays and
inserting text (which slows down with more overlays) dominates.
However, the time required to display the first page of results is
generally more important to the user experience.
This provides the same interface as the streaming JSON parser, but
reads S-expressions incrementally. The only difference is that the
`notmuch-sexp-parse-partial-list' helper does not handle interleaved
error messages (since we now have the ability to separate these out at
the invocation level), so it no longer takes an error function and
does not need to do the horrible resynchronization that the JSON
parser had to.
Some implementation improvements have been made over the JSON parser.
This uses a vector instead of a list for the parser data structure,
since this allows faster access to elements (and modern versions of
Emacs handle storage of small vectors efficiently). Private functions
follow the "prefix--name" convention. And the implementation is much
simpler overall because S-expressions are much easier to parse.
Previously, search started the async notmuch process directly. Now,
it uses `notmuch-start-notmuch'. This simplifies the process sentinel
a bit and means that we no longer have to worry about errors
interleaved with the JSON output.
We also update the tests of Emacs error handling, since the error
output is now separated from the search results buffer.
This provides a new notmuch-lib utility to start an asynchronous
notmuch process that handles redirecting of stderr and checking of the
exit status. This is similar to `notmuch-call-notmuch-json', but for
asynchronous processes (and it leaves output processing to the
caller).
Since the part commands are no longer tied to a button, but can be
applied with point anywhere within a part, bind the part commands
keymap to "." everywhere in the show buffer. This lets you save or
view parts without having to navigate to the part button, and is
particularly useful for parts that have no button.
This removes the un-prefixed MIME part commands from the part button
keymap, but that's okay because those clashed in annoying ways with
show buffer bindings like "s" for search. RET on part buttons is
unaffected, which is the most important part button binding.
This unifies the part button actions and the underlying part action
functions into single interactive command that simply applies to the
part containing point using the just-added part p-list text property
instead of button properties. Since all part actions can be performed
by applying the appropriate mm function to an mm-handle, this patch
abstracts out the creation of mm handles, making the implementations
of the part commands trivial. This also eliminates our special
handling for part save in favor of using the appropriate mm function.
This necessarily modifies the way we handle the default part button
action, but in a way that does not change the meaning of the
notmuch-show-part-button-default-action defcustom.
Since these commands are no longer specific to buttons, this patch
eliminates the extra metadata stored with each button. This also
eliminates one rather special-purpose macro for a collection of
general purpose part handling utilities.
This is similar to what we already do with the message p-list, though
we apply the part's text property to the whole part's text, in
contrast with the message p-list, which is (rather obscurely) only
applied to the first character.
Previously, we lost any text properties applied to part buttons or
wash buttons when they were toggled because `insert' directly copies
the text properties of the string being inserted. Fix this by
capturing the properties applied to the button beforehand and
re-applying them after inserting the new text.
For such a simple regexp, this was broken in a very complicated way.
The intent was to strip the newline (and potentially other whitespace)
off the end of the error string so there wasn't an extra newline in
the error signal. However, the regexp was deeply dependent on the
active syntax table and the subtleties of $. We didn't notice this
because all notmuch major modes put ?\n in the whitespace class, which
makes this behaved as intended: the "\\s " matches all newlines, but
by matching the newline character, causes the $ *not* to match
*except* where it matched the empty string at the very end of the
string, which was not followed by a newline.
However, if the syntax table declares ?\n to be non-whitespace
(lisp-mode declares it as endcomment, and is likely to be the mode
you're in when testing functions), then this regexp behaves completely
differently, matching trailing spaces at the end of every line within
the string.
The solution is to say what we mean for whitespace *and* to switch
from $ to \', which matches only the end of the string, rather than
the end of each line. Both are necessary or this will strip away
interior newlines, which is not what we want.
