| Commit message (Collapse) | Author | Age | Files | Lines |
|---|
| |
|
|
|
| |
The Berkeley Mono code ladder is now Coding XS (11pt), Coding M (13pt),
Coding L (14pt), and Coding XL (16pt).
|
| |
|
|
|
|
|
|
|
|
|
|
| |
M-SPC now steps only through attached agents (the ones with a live
on-screen buffer), so the fast chord stays among what's already up.
M-S-SPC steps through all active agents and attaches a detached tmux
session when it lands on one, which is what M-SPC used to do.
I moved the shared step logic into cj/--ai-term-step-among and added
cj/--ai-term-attached-agent-dirs to filter the queue to attached
buffers. I bound both chords globally and in eat-semi-char-mode-map so
they work from inside an agent buffer. C-; a n keeps cycling all.
|
| | |
|
| |
|
|
|
|
| |
I replaced the mixed Fontaine presets with seven named profiles that persist across restarts. The mode line, echo area, and minibuffer stay in Berkeley Mono.
Nov applies the shared Reading profile inside each EPUB buffer without changing the global selection.
|
| |
|
|
|
|
|
|
| |
- dirvish-config, dwim-shell-config: both call cj/xdg-open (external-open) but relied on init.el requiring it before them, and dirvish's require comment misattributed it to system-utils. Added the require to both and corrected the comment and both Runtime-requires headers.
- tramp-config: revert-without-query was set to '(".*") inside the TRAMP module, so every file in Emacs reverted without confirmation. Scoped it to the TRAMP path regexp, so local files keep their revert prompt and remote files skip the round-trip.
- tramp-config: removed the sshfast method (referenced nowhere, scp and sshx handle everything) and a duplicate tramp-own-remote-path add-to-list.
Tests: new runtime-requires assertions for both modules, red first under make test-file. The tramp :config edits are verified in the daemon. Suite green.
|
| |
|
|
|
|
|
|
|
|
|
|
| |
- vc-config: C-; v d matched literal +/- source text via consult-line, not gutter hunks. It now completes over git-gutter's hunk list and jumps to the chosen line.
- vc-config: clipboard clone ran git synchronously, freezing every frame for the whole clone. It now runs async with a sentinel that opens the clone on success and surfaces the process buffer on failure.
- vc-config: magit-blame bound D and S to the same command. D is now difftastic-magit-diff, matching the transient.
- vc-config: dropped the :commands autoload for git-timemachine-show-selected-revision, a function the package never defined. The phantom appeared in M-x and errored.
- dev-fkeys: both F4 chained-compile handlers armed a hook on the global compilation-finish-functions before their compile ran, so a quit left it live and the next unrelated compile fired the chain. The one-shot hook now installs buffer-locally in the compilation buffer, same shape as the projectile cache-revert hooks.
- test-runner: outside a project with no global test directory, three commands crashed with wrong-type-argument on a nil path. Discovery now returns nil and the commands signal user-error.
- diff-config: removed the global "-w" ediff default, which made every session ignore whitespace (indentation-only changes compared as identical). Whitespace-ignore stays available as ediff's per-session toggle.
- restclient-config: the C-; R bindings went through raw global-set-key, silently depending on keybindings.el loading first. They now use a prefix keymap registered like the other C-; prefixes.
- httpd-config: simple-httpd loaded on a 1s timer and created www/ on every startup. It now defers until impatient-mode needs it, and the doc root is created at package load.
|
| | |
|
| |
|
|
|
|
| |
Classic go-mode and js-mode buffers got none of the ts-mode setup, and web-mode got the format key without the promised company/flyspell/LSP. All three now run the shared setup. The web-mode LSP attach guards on the HTML language server, so machines without it stay silent instead of prompting.
gopls, clangd, clang-format, bash-language-server, shfmt, and shellcheck now warn at load when missing. The shfmt, shellcheck, and clang-format blocks gate on :if, which evaluates once at startup, so an absent tool silently disabled that setup until the next restart. I also put ~/go/bin on exec-path at load so the gopls check doesn't misreport a gopls installed there.
|
| |
|
|
|
| |
- cj/media--play-sentinel returns the per-process closure.
- Three new tests pin the reap-on-exit behavior.
|
| |
|
|
|
|
|
|
| |
- open-this-file-with splits the typed command and calls call-process.
- media-play-it resolves streams via a yt-dlp capture, then start-process.
- Paths, URLs, and player args never meet a shell.
- yt-dlp stderr is captured separately so warnings can't parse as URLs.
- Stream resolution now blocks briefly. mpv (the default) is unaffected.
|
| |
|
|
|
|
|
| |
- D in the playlist buffer deletes a saved playlist.
- Candidates span both m3u sources: local library and MPD dir.
- A typed yes/no confirms deletion.
- Deleting the loaded playlist's file clears its association.
|
| | |
|
| |
|
|
| |
marginalia-align has been right for months, but annotations from the custom completion categories (radio stations, music files, and everything built through the system-lib table helpers) never went through marginalia. They rendered unaligned. A registration helper adds a builtin registry entry per category at table construction, which routes the table's own annotation function through marginalia's aligned field. The radio table converts from an affixation function with hand-rolled padding to a plain annotation, since marginalia now owns the alignment. Self-maintaining: any future custom category built through the helpers registers itself.
|
| |
|
|
| |
The header block (playlist name, now-playing hero, key legend) is an overlay string anchored at buffer position 1, so scrolling the list carried it off the top, key legend included. A handler on the buffer-local window-scroll-functions re-anchors the overlay at the window's display start on every scroll, so the block stays pinned at the top of the window. The header refresh timer now anchors at the same position instead of the buffer top, so the once-a-second update can't yank a scrolled header back.
|
| |
|
|
|
|
| |
The playlist header is a multi-line overlay string anchored at buffer position 1. With point on the top row, next-line's visual motion stepped through the header's screen lines. Those all map back to position 1, so arrows never moved point. The buffer now moves by logical lines (each row is one), which skips the header display entirely.
The same shared anchor put row 1's number above the header instead of beside its track: both overlay strings sit at position 1 and the number rendered first. Number overlays now outrank the header's priority, so the number lands next to its own row.
|
| |
|
|
| |
Emacs draws the filled cursor only in the selected window, and the playlist dock is glanced at from other windows constantly. The pinned cursor was invisible exactly when it was wanted. The current row's number now renders in inverse video, updated as point moves and re-applied after each renumber. The mark shows whether or not the window has focus, and the row band from hl-line was already sticky.
|
| | |
|
| |
|
|
|
|
| |
With point pinned at the row start, redisplay drew the cursor after the number's before-string, on the album-art thumbnail, where a block cursor is invisible. A cursor property on the number string makes redisplay draw it on the number.
The dock also kept reopening too short: its height memory remembered whatever size the window had at toggle-off, including squeezes from window churn. The squeezed fraction then overrode the default on every later open. The default rises to a third of the frame, and the toggle discards captured heights below it (a deliberate enlargement still sticks for the session).
|
| |
|
|
| |
Vertical motion over thumbnail rows and the stretch-space that right-aligns the metadata drifts point to arbitrary visual columns, usually line end. The rows aren't editable text, so the cursor now lives at the row start (the number column). A buffer-local post-command snap sends every landing back there. Kill, shift (C-up/C-down), and play already act on the whole row, so all track operations happen from the number. An active isearch keeps point until it ends, then the snap lands on the match's row.
|
| |
|
|
|
|
| |
Opening the playlist used to jump to EMMS's selected track, which stays set while stopped, so the view opened deep in the list with point mid-row. It now lands on the playing track's row (beginning of line, upper third of the window) when a song is playing, and at the top of the list when stopped. Both entry points share the logic, and the playlist buffer gets hl-line so the current row stays findable on album art.
Also fixed: directory adds handed the raw tree to emms-add-directory-tree, which adds every file it finds. Cover art and liner notes became playlist rows. All three add paths now route through a filtered walk that keeps only accepted music extensions. The m3u loader applies the same filter to local paths (stream URLs pass through), so old playlists with saved cover lines stop re-importing them.
|
| |
|
|
|
|
|
|
|
|
| |
The radio tag search now completes over popular tags from radio-browser's /json/tags endpoint (station-count order, capped at 500, cached per session), so the prompt offers real tags instead of a blind guess. Free-form input still works, and a failed fetch degrades to plain input.
Station listings align into columns: the votes field pads to a fixed width, and each annotation pads out to the widest candidate name.
Playlist rows carry a numeric overlay prefix, rebuilt after every change by a debounced after-change hook, so the cursor stays visible on cover-art thumbnails and each row's position is readable. Overlays leave the buffer text to EMMS.
Also fixed: the search now trims queries before they hit the API (a trailing space used to reach it as %20 and match nothing).
|
| |
|
|
| |
The docstring promised toggle-off puts the displaced buffer back via quit-restore-window. Toggle-off deliberately deletes the window instead: the slot's quit-restore parameter goes stale when several agents share it. That staleness is the "F9 shows another agent" bug the delete-window design killed.
|
| |
|
|
| |
ghostel retargets default-directory via OSC 7 on every cd in the agent shell, so close (and quit with no project argument) computed the wrong aiv- session name after a cd. The kill then missed the real session and orphaned the agent, or hit a different project's session. The buffer name ("agent [basename]") never changes, so teardown now derives the basename from it.
|
| |
|
|
|
|
|
|
| |
The expander ignored BYDAY on monthly and yearly rules and stepped DTSTART's day-of-month instead, so a "2nd Wednesday" series rendered on the 12th of every month, the wrong weekday most months. It now resolves nth-weekday entries (2WE, -1TU), bare weekdays with BYSETPOS, and yearly BYMONTH+BYDAY, so each occurrence lands on the day the rule names.
Nothing read the VEVENT STATUS property, so cancelled events rendered as normal meetings: a cancelled instance of a series reappeared at its original time, and standalone cancelled events stayed on the agenda. parse-event now drops STATUS:CANCELLED events, which also kills cancelled series masters. A cancelled RECURRENCE-ID override removes its occurrence instead of overriding it.
The fixes are coupled: until BYDAY generates the right dates, a cancelled override can't match the occurrence it removes.
|
| |
|
|
|
|
| |
The agenda file-list cache held a 1h TTL, so once an hour the next agenda open re-walked projects-dir from scratch. That cache tracks only which files exist, never their contents, so task edits and completions are never staled by it. The only thing a long TTL delays is a new project's todo.org appearing. That's rare, so I raised it to 24h.
For that case I bound cj/org-agenda-refresh-files (the manual force-rescan) to S-<f8>, the free slot in the F8 agenda family (<f8>, s-<f8>, C-<f8>, M-<f8>).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
format-timestamp built its date from the start and took only the hour and
minute from the end, so the end date was thrown away. A conference running Jul
20 09:00 to Jul 23 17:00 produced <2026-07-20 Mon 09:00-17:00>, byte for byte
the same timestamp as a same-day meeting, and the agenda showed it on the first
day only. All-day spans collapsed the same way, to a bare <2026-07-20 Mon>.
An event whose last day is later than its start now renders as an org range,
<start>--<end>, which org parses as an active-range and shows on every day it
covers.
DTEND is the non-inclusive end of the event (RFC 5545 3.6.1), so an all-day
event's last day is DTEND-1. Getting that backwards would have been worse than
the bug: a one-day all-day event carries DTEND = start+1, so a naive range
would turn every single all-day event into a two-day one. The decrement only
applies when both ends are date-only. A date-only start with a timed end is
malformed, and treating it as all-day would put the last day before the start
and emit a backwards range. That case falls through to the same-day form, as
before.
I split out format-stamp and format-hhmm so both halves of a range and the
compact same-day form build from one place.
Same-day events are unchanged, pinned by two tests: the compact HH:MM-HH:MM
form, and a single all-day event staying a single stamp.
One unrelated bug stays open. A timed event with no DTEND renders as <date> and
drops its time, so a 09:00 meeting reads as all-day. I left it alone.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
RFC 5545 3.3.10 bounds a recurrence inclusively: when UNTIL lines up with the
recurrence, that date is the last instance. The expansion loops guarded on
calendar-sync--before-date-p, a strict comparison, so the instance landing
exactly on UNTIL was dropped. Every bounded series silently lost its last
meeting from the agenda. A series whose UNTIL equals its start date lost the
only instance it had and vanished.
I added calendar-sync--date-on-or-before-p beside --before-date-p and swapped
the three UNTIL sites: the simple-recurrence loop feeding daily, monthly and
yearly, plus both weekly checks. --before-date-p is unchanged, since a date is
still not before itself.
The two UNTIL property tests asserted the wrong invariant. They required every
occurrence to fall strictly before UNTIL, the exclusive reading the RFC
contradicts, so they pinned the defect they should have caught. The property is
now on-or-before. An upper bound alone can't catch a dropped occurrence, so I
added one asserting the series reaches its UNTIL date.
Reverting the fix failed three daily tests and one property test, and no weekly
ones. The weekly loop was fixed but unguarded, so it has its own test now.
The comparison stays date-granular, matching --date-to-time. An UNTIL carrying a
time of day earlier than the event's start will include that final instance
where the RFC would exclude it. Making it datetime-precise means reworking
--date-to-time and the timezone conversions feeding it, a much larger change
than this bug warrants.
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| |
cj/timeformat was defvar'd inside cj/add-timestamp-to-org-entry, so the symbol
wasn't special until the command had run once. Until then a let around the call
bound it lexically and never reached the function. I moved it to top level with
a docstring. It has no other callers.
The nested defvar also poisons tests. ERT runs alphabetically, so an earlier
test that calls the command makes the symbol special retroactively. A later
special-variable-p check then passes on test ordering rather than on the code,
reading green in a full-file run and red in isolation. The new guard snapshots
special-variable-p into a defconst at load, before any test body runs.
I filled in the missing coverage: point and the following line survive the
insert, empty and unicode time strings, an empty buffer, and a read-only buffer
signalling rather than dropping the stamp. The empty-string case pins a trailing
space (current behavior, characterized rather than changed).
|
| |
|
|
| |
One command flips markdown/gfm buffers to their view-mode partner and back. Checks the most-derived mode first; gfm-view derives from gfm-mode, so the dispatch resolves correctly for all four modes.
|
| |
|
|
| |
The phone's FTP server can't open active-mode data connections back to the client, so LIST hangs under ange-ftp's default. Passive mode fixes it and is the safer default for any NAT'd server. Requires a CLI ftp client (GNU inetutils on Arch).
|
| |
|
|
|
|
|
|
| |
- A failed wf-recorder start still writes a ~0.5s stub .mkv.
- Stubs littered the recordings directory and matched *.mkv globs downstream.
- The start path now stamps the output path on the process.
- The sentinel's failed-start branch deletes that file.
- User stops and real recordings never trigger the delete.
|
| |
|
|
| |
Agents drive Signal via signal-cli and signal-mcp, so the interactive client earns no keep. I moved signal-config.el and its seven test files to archive/ per the pasture convention, dropped the require from init.el, removed the dashboard's Signal launcher (row sizes now 5 4 3 2, tests updated), and unregistered the C-; M prefix from the running daemon. The ~/code/signel fork repo is untouched. The spec record stays IMPLEMENTED with a retirement history line, and the messenger-unification draft carries a premise-shift note.
|
| |
|
|
|
|
| |
lsp-pyright's :hook lambda called require and lsp-deferred unguarded on every python-ts buffer, so a machine without pyright still got the LSP attach prompt that cj/python-setup's guard exists to prevent. The guarded branch now owns the require and the attach, and the hook is gone.
Classic python-mode now runs the same setup hooks as python-ts-mode, so treesit-auto's grammar-unavailable fallback keeps indentation, keys, and LSP. Four tests cover the guard branches and pin that every python hook entry stays a named function.
|
| |
|
|
| |
The :custom block set checkdoc-arguments, a variable that doesn't exist. Checkdoc has no such option and this flycheck runs checkdoc through a fixed subprocess form, so the intended suppression of two warning types never happened. Removing it changes no behavior. If those checkdoc warnings need silencing, that's a feature against flycheck-emacs-lisp-checkdoc-form.
|
| |
|
|
| |
:after (flycheck package-lint) waits for both features, but nothing loads package-lint on its own (it's flycheck-package's dependency, pulled in by the require), so the setup never ran and elisp buffers never got the package-lint checker. Keying on flycheck alone loads the chain.
|
| |
|
|
| |
C-c C-q was bound via :bind (:map json-ts-mode-map) inside the jq-mode use-package, which defers the binding to jq-mode's load. Nothing ever loaded jq-mode, so the key was dead. The binding now keys on json-ts-mode's load, and jq-interactively's autoload pulls jq-mode in on first press.
|
| |
|
|
| |
leetcode's :custom set url-debug t, a debugging leftover that turned on url.el request logging for the entire session the moment leetcode loaded. I also declared code-dir for the compiler.
|
| |
|
|
| |
Incoming Telegram messages were invisible unless their buffer was on screen; telega ships notifications but leaves the mode off. telega-notifications-mode now turns on in the package's :config, so it activates when telega loads and respects telega's own per-chat mutes. Routing through a shared messenger notifier stays with the unification task.
|
| |
|
|
| |
The script-absent fallback called notifications-notify, which isn't autoloaded, so a fresh session without the notify script would die on void-function instead of showing the toast. The fallback now requires notifications.el when the function is unbound. Found while giving slack-config the same fallback.
|
| |
|
|
|
|
| |
cj/slack-notify passed raw message bodies to a bare start-process. A multiline message made a giant toast, and a machine without the notify script errored inside the condition-case, so the notification silently vanished. It now routes through the same shape signel uses: whitespace collapsed and the body truncated at 120 chars, sound gated by a new cj/slack-notify-sound defcustom (silent by default), and a notifications-notify fallback when the script is absent. The fallback requires notifications.el first, which isn't autoloaded. cj/slack-test-notify exercises the same delivery path.
The shared cj/messenger-notify extraction that collapses the signel and slack copies stays with the messenger-unification task. Ten tests cover the formatter, the delivery routing, and the DM/mention predicate wiring the audit called untested.
|
| |
|
|
|
|
|
|
| |
Two picker-cache bugs from the config audit, verified against current code first.
The cache docstring promised clearing on signel-stop, but nothing cleared it, so a stale contact list survived a relink or reconnect. I advised a named clear function :after signel-stop, which makes the docstring true.
An empty listContacts result cached nil, indistinguishable from a cold cache, so a zero-contact account re-ran the picker's blocking fetch (up to the 3s timeout) on every open. The fetch now caches an empty sentinel and the picker reads through cj/signel--cached-contacts. One existing test pinned the nil behavior and now asserts the sentinel instead.
|
| |
|
|
|
|
|
|
|
|
| |
A fresh agent session now asks which runtime to run: claude, codex, or any local ollama model via codex --oss. RET keeps claude, so the common launch stays Enter-Enter. Reattaches and redisplays never ask. The picker fires only when neither a live agent buffer nor a surviving tmux session exists.
I mirrored the runtime names and launch strings from the bin/ai launcher so the two stay one mental model, and the choice list itself comes from "ai --print-runtimes" when that launcher is installed, reusing its live ollama scan and dead-server timeout. A static claude-first list stands in without it. The explicit --local-provider=ollama flag is deliberate: configuring the provider through config.toml silently does nothing.
The chosen command threads through show-or-create into the tmux launch line. On reattach, tmux new-session -A ignores it, so no reattach path changes behavior. cj/ai-term-agent-command keeps its meaning as the claude runtime's full line.
Thirteen tests cover the runtime-to-command map, the print-runtimes parsing, the launcher fallback, and the launch-command override.
|
| |
|
|
| |
mu4e 1.14 calls mu4e--compose-setup-completion directly during compose setup, never via mu4e-compose-mode-hook, so the activate path's remove-hook did nothing and the deactivate path's add-hook installed a function that was never hook-driven. The mu4e-compose-complete-addresses variable already gates the built-in completion on both paths. I also declared that variable for the compiler, which clears the module's two free-variable warnings.
|
| |
|
|
|
|
| |
The selection buffer captures MIME handles when it opens, but viewing another message kills the buffer a handle points at. Saving through such a handle died deep in mm-decode, or could write another message's bytes. The save path now signals a user-error naming the attachment when its handle's buffer is dead, and it checks before any mu4e MIME support loads, in the same spot as the existing no-handle check.
I added tests for the stale-handle error, the healthy live-handle save, and the mid-batch failure behavior (the error propagates and files already written stay).
|
| |
|
|
|
|
|
|
|
|
|
|
|
|
| |
Seven config-audit findings in mail-config, each verified against the installed mu4e 1.14 source before changing anything:
- The C-; e account nav commands now require mu4e before searching. The maps register eagerly at startup, but mu4e-search has no autoload cookie, so a nav key pressed before mu4e's first launch signaled void-function.
- With msmtp absent, both send variables now get cj/mail--send-mail-unavailable, which signals a user-error naming msmtp. The old fallback left message-send-mail-function nil (the top-level defvar pre-empts message.el's default), so the first send died with "invalid function: nil".
- I removed the unconditional org-msg-edit-mode advice on the two reply commands. org-msg-post-setup already runs on mu4e-compose-mode-hook and applies org-msg-default-alternatives itself, so the advice forced org-msg onto text-only replies and re-ran a major mode org-msg had already set up.
- I dropped the save-attachment headers action: mu4e-view-save-attachments takes no message argument and reads MIME parts from the view buffer, so it never worked from headers.
- I deleted the obsolete HTML view knobs (mu4e-view-prefer-html and both mu4e-html2text-command sets): the shr-based view has ignored them since mu4e 1.7.
- I reordered the contexts so cjennings.net comes first: with pick-first, gmail-first made gmail the silent default account for the first compose.
- I dropped mu4e-starred-folder (not a mu4e variable, never had an effect) and the obsolete mu4e-maildir alias.
New tests cover the nav-command mu4e load and the msmtp-absent fallback.
|
| |
|
|
|
|
| |
I removed the remove-hook shim that swapped the retired popup handler out of a live daemon on module reload. The handler it names was deleted long ago and both daemons have restarted since, so the shim's only remaining effect was a not-known-to-be-defined warning.
I also declared org-capture-mode-map for the compiler (it's only ever touched inside with-eval-after-load) and reflowed one over-wide docstring. The module now byte-compiles with zero warnings.
|
| |
|
|
|
|
| |
The capture popup frame opens showing the daemon's last buffer, and only the capture UI or the reap-on-finalize hook clears it. When a capture aborts before its UI paints (a C-g, an erroring template, a path that skips cj/quick-capture), the popup lingers on whatever was current. If that was a live eat/vterm terminal, eat sizes the terminal to the small popup window and clamps the real frame to the popup's rows.
Two guards keep the popup from ever holding a size-sensitive live buffer. One on after-make-frame-functions neutralizes the frame at creation, the root cause. One on window-buffer-change-functions catches the abort and stray-switch paths the finalize reap misses. Both repoint any non-capture-UI window to *scratch*, sparing *Org Select* and CAPTURE-* buffers so a live capture is never disturbed.
|
| |
|
|
|
|
|
|
| |
tmux 3.7b won't emit sixel until it learns the client's cell pixel size via CSI 14/16/18 t. EAT 0.9.4's parser has no CSI t clause and drops the queries, so images never render. I added a :before advice on eat--t-handle-output that scans the raw output and answers through the terminal's own input function (eat--t-with-env binds eat--t-term). I verified it live: images render and survive window switches, scrolling, and resizing.
I guarded the advice-add so it installs only while EAT itself can't answer. Two answerers give tmux a double reply, and it forwards the second one's raw bytes into the pane as keystrokes. The guard keys on eat--t-send-window-size-report, the function an upstream clause would define.
I kept an upstream-shaped parser-clause patch locally for a PR to akib/emacs-eat.
|
| |
|
|
| |
calendar-sync--get-exdates read (match-end 0) after split-string, whose internal matching clobbers the match data. The scan position jumped back to a comma offset inside the value and re-matched the same EXDATE line forever, growing the result list until the kernel OOM killer took the whole session down (twice, 2026-07-13). The fix captures the line's end position before the split. The comma-separated tests added with a9e61207 were the first code to hit this path. All 12 exdates tests and all 56 calendar-sync test files now pass in milliseconds.
|
| | |
|
