The performance notes documented one-off benchmarking of the switcher re-rank
and the fixes already applied; it is no longer useful as living documentation.
Drop docs/PERF.md and the references to it in README (layout tree), AGENTS.md
(reference-docs list and the build-type note), and the CMakeLists comment.
- docs/radials.example.json: the config comments and the launcher-door action
names were still Ollie-era — 'Copy to .../olliepalette/radials.json' ->
'deft/radials.json', and ollie_goto_file/ollie_goto_symbol/ollie_command_palette
-> deft_goto_file/deft_goto_symbol/deft_command_palette (the objectNames the
radial plugin actually resolves; verified against radialconfig.cpp).
- docs/PLUMBING.md: replaced the dead 'See PLAN.md' pointer with the real
implementation location (src/plumb/, plumb_lib + plumbplugin).
- README: link docs/radials.example.json from the radial Configure note.
- AGENTS.md: drop the environment-specific LSP-bridge remark (not universal
contributor guidance); keep the plain 'read before you edit' rule.
PERF.md and SAM.md were already accurate and are unchanged.
Remove docs/PLAN.md (goal/constraints/milestone roadmap — stale planning prose)
and the README's Status/milestone section; the README already documents what
the plugins are and do. Refresh the layout tree's docs/ listing accordingly.
Add AGENTS.md: a working guide for contributors/agents — the one-.so-per-feature
architecture and shared libs, build/test/install commands, naming and style
conventions, the two load-bearing design decisions (eventFilter'd launcher keys
vs ambiguous shortcuts; cross-plugin calls by action objectName with no link
dependency), and the gotchas (osort collision, per-plugin frecency files,
live-Kate behaviour being unverifiable in CI).
Add the docs/img/ screenshots and embed one in each plugin's subsection:
palette (deft-mx), radial (deft-radial), acme (deft-acme-select), sam
(deft-sam), and plumb (deft-ollie-plumb). The switch plugin has no screenshot
yet.
A home-prefix install puts the plugins outside Qt's default search path, so
Kate would not list them until QT_PLUGIN_PATH includes the install dir — the
step a bare 'cmake --install' silently left undone. Solve it the way the ollie
project does (a Plasma login env snippet), and automate the common case:
- Add 99-deft.sh.in, a Plasma env snippet that prepends the install's qt6
plugin dir to QT_PLUGIN_PATH. configure_file() fills in the absolute path.
- On a user-local install (prefix inside $HOME), install(CODE) drops the
configured 99-deft.sh into ~/.config/plasma-workspace/env/ so the lookup
path is wired up with no manual step; Plasma sources it at next login. The
installer also prints a one-line export to use the plugins immediately.
- On a system-prefix install it does not touch the user's config; it just
prints where the plugins went and leaves the ready snippet in the build dir.
README: add a 'Making Kate find a user-local install' section explaining that
QT_PLUGIN_PATH must include ~/.local, that the installer handles it, and how to
activate it in the current session (or a non-Plasma shell).
Verified: cmake --install with the default prefix installs the six plugins and
writes ~/.config/plasma-workspace/env/99-deft.sh with the correct path.
cmake --install previously defaulted to CMAKE_INSTALL_PREFIX=/usr/local (needs
root) and, via KDE_INSTALL_PLUGINDIR, to lib/<arch>/plugins/ — a directory Kate
does NOT scan (it scans lib/<arch>/qt6/plugins/, per QT_PLUGIN_PATH). So a plain
install both failed on permissions and would have put the plugins where they
can't be found.
Follow the sibling ollie project's approach instead of hand-rolling path logic:
- Default CMAKE_INSTALL_PREFIX to $HOME/.local when the user did not pass one.
ECM's KDECMakeSettings resets an unspecified prefix back to /usr/local, so we
capture CMAKE_INSTALL_PREFIX_INITIALIZED_TO_DEFAULT *before* including it and
re-apply the user-local default *after*, where it is not clobbered.
- Point KDE_INSTALL_PLUGINDIR at QT6_INSTALL_PLUGINS (prefix-relative
lib/<arch>/qt6/plugins) so kcoreaddons_add_plugin(INSTALL_NAMESPACE
kf6/ktexteditor) lands where Kate looks. Set the *normal* variable (the macro
reads it directly; a cache-only FORCE would be shadowed) as well as the cache.
Result: `cmake --install build` with no flags installs all six plugins to
~/.local/lib/<arch>/qt6/plugins/kf6/ktexteditor/ with no root.
README: replace the manual install -D loop with `cmake --install build`, state
clearly that it installs to the home dir by default, show the resulting path,
and document --prefix /usr for a system-wide install.
Several plugins reshape or suppress Kate's defaults (radial takes the right
mouse button, acme rebinds Ctrl+A/E/H/U/W + double-click + mouse chords and
kills X11 primary paste, palette/switch/plumb claim Alt+X/P/G and F2). Add a
callout at the top of 'The plugins' so users enable them deliberately rather
than being surprised.
Add a 'The plugins' section with a short description and a defaults-based usage
guide for each of the six plugins: palette (M-x + the :-verb table + frecency),
switch (Go to File/Symbol, ctags, project scoping, cache), radial (open/choose
gestures + radials.json config), acme (line keys, pair select, mouse chords),
sam (tool view, Run, X/Y multi-file), and plumb (F2 send, edit port, fallback).
Trim the now-redundant usage prose from the install section and point it at the
new section.
The transient core classes were already deleted in the split, but the shared
command library and a scattering of comments/ids still carried the old project
name. Finish de-Ollie-ing the tree so nothing but 'Deft' remains:
- src/commands: OllieCommands -> DeftCommands; files olliecommands.{h,cpp} ->
deftcommands.{h,cpp}; lib target olliecommands_lib -> deftcommands_lib; tests
test_olliecommands_live/test_ollieregistration -> test_deftcommands_live/
test_deftregistration (ctest names deftcommands_live/deftregistration).
palette plugin updated to match (include, fwd-decl, member type, ctor call,
runVerb call, CMake link).
- Config paths: radial's user config moves olliepalette/radials.json ->
deft/radials.json, matching the switch plugin's earlier olliepalette/ -> deft/
move; comments updated.
- Stale ollie: id prefixes in the project benchmark (ollie:file:/ollie:sym: ->
deft:file:/deft:sym:) and sample paths in test_palettemodel referencing the
deleted src/plugin/ollieplugin.cpp.
- Comments still naming OllieView/OlliePlugin corrected to the real post-split
classes (SamView, PlumbView) or to deft paths.
- README layout refreshed (deftcommands.{h,cpp}).
PLAN.md keeps the Ollie names inside its pre-split historical prose, as flagged
by the architecture note at its top.
Full clean rebuild + 18/18 ctest green; palette and radial reinstalled.
README: replace the single-deft.so install instructions with the per-feature
plugin table (palette, switch, radial, acme, sam, plumb), a glob install loop,
and a note that the plugins cooperate by resolving each other's actions by
object name; refresh the layout tree (plugin/ gone, switch/ added, palette/ and
radial/ now hold their plugins) and mark milestones 5 and 6 DONE.
PLAN.md: add an architecture note at the top flagging the split and that the
design prose below refers to the former monolith as historical.
Stage 3c, the final split. The remaining core OllieView carried two distinct
features plus the ":"-verb pack; each becomes its own KTextEditor plugin and
the monolithic core 'deft' plugin is removed entirely.
deft_palette.so ([deft:util] palette): new src/palette/paletteplugin.{h,cpp}
(PalettePlugin/PaletteView) + paletteplugin.json. Owns the command palette
(Alt+X / M-x): collectActions over the window's GUI clients, the ":"-verb
pack (OllieCommands, constructed once in PalettePlugin), positionPalette,
frecency ranking, and the runAction dispatcher. Its "Go to File/Symbol"
entries no longer call switcher code directly — they trigger the switch
plugin's window actions deft_goto_file / deft_goto_symbol by objectName via
findActionByName, so no cross-plugin link is required. Links palette +
olliecommands_lib.
deft_switch.so ([deft:util] switch): new src/switch/switchplugin.{h,cpp}
(SwitchPlugin/SwitchView) + switchplugin.json. Owns the project file (Alt+P)
and symbol (Alt+G) switchers: ProjectIndexer, KateProjectBridge (follows
kateprojectplugin, invalidates the index on project change), async
cache-populated palettes, ctags 1-based->0-based symbol jumps. Registers
deft_goto_file / deft_goto_symbol as window actions so palette and radial can
reach them. Links palette + project_lib.
findActionByName now also searches window()->actions(), since sibling doors are
registered via QWidget::addAction (no XMLGUI client); this is what makes the
cross-plugin door resolution actually work.
Each plugin keeps its own FrecencyStore in a separate file under deft/
(palette.frecency.json, switch.frecency.json) so the two independent rankings
can no longer clobber one another through a shared file; the old config dir
olliepalette/ and the ollie: id prefixes are retired in favour of deft/ and
deft:cmd:/deft:file:/deft:sym:.
Deletes src/plugin/ (ollieplugin.{h,cpp,json}, CMakeLists) and the core 'deft'
CMake target; src/CMakeLists.txt now builds acme, commands, switch under the
KF6TextEditor guard and no longer adds the plugin subdir. The suite is now six
one-feature plugins: palette, switch, radial, acme, sam, plumb.
Stage 3b of the split. acme becomes its own KTextEditor plugin: new
src/acme/acmeplugin.{h,cpp} (AcmePlugin/AcmeView) + acmeplugin.json
("[deft:nineify] acme"), built as deft_acme.so with its own event filter:
the ^A/^E/^H/^U/^W line keys, double-click bracket/quote pair selection, and
the left-button mouse chords (left+middle=Cut, left+right=Paste) plus X11
primary-paste suppression. The bracket/quote/step helpers move here too.
The core deft plugin drops all acme code; its event filter now only
intercepts the Alt+X/P/G launcher-door keys for palette + switch.
Stage 3a of the util split. radial becomes its own KTextEditor plugin:
new src/radial/radialplugin.{h,cpp} (RadialPlugin/RadialView) + radialplugin.json
("[deft:util] radial"), built as deft_radial.so with its own event filter
(mouse-button gesture + context-menu suppression + pane-focus pointer warp).
The optional Wayland pointer-warp glue (waylandcursorwarp.* + the ECM
wayland-scanner protocol lib) moves from src/plugin into src/radial, since the
warp is only used to follow focus after a radial pane action.
Door object names renamed ollie_* -> deft_* (deft_command_palette /
deft_goto_file / deft_goto_symbol) so the radial resolves the sibling plugins'
launcher actions via the GUI factory; built-in radial config + test updated.
The core deft plugin drops radial code, members, the radial branches of its
event filter, and the Wayland glue; it keeps palette + switch + acme keys.
Stage 2b of the multi-.so split. plumb becomes its own KTextEditor plugin:
new src/plumb/plumbplugin.{h,cpp} (PlumbPlugin/PlumbView) + plumbplugin.json
("[deft:nineify] plumb"), built as deft_plumb.so. The F2 plumb action, the
edit-port reader wiring, plumbAtCursor/onPlumbEdit, and the acme isFileChar/
fileTokenAt file-token helpers move out of OllieView into PlumbView.
The core deft plugin no longer links plumb_lib, owns the Plumber, or carries
the plumb/file-token code. activeDocumentDirFor stays in the core (the
switchers still use it) and is duplicated into the plumb plugin.
Stage 2a of the multi-.so split. sam becomes its own KTextEditor plugin:
new src/sam/samplugin.{h,cpp} (SamPlugin/SamView) + samplugin.json
("[deft:nineify] sam"), built as deft_sam.so. The sam tool-view wiring and
the runSamProgram/applySamToDocument/runSamFileLoop glue move out of
OllieView into SamView; sampanel.{h,cpp} move from src/plugin to src/sam.
The core deft plugin no longer links sam_lib or owns the sam panel.
Rename the C++ namespace katecustom -> deft and include guards
KATECUSTOM_* -> DEFT_* across all sources. Rename the plugin artifact
olliepalette.so -> deft.so (CMake target 'deft') and set the plugin
display name to 'Deft'. Update README title/install paths.
File names, class names (Ollie*), and the src/ directory grouping are
left for the follow-up stages that split features into per-plugin .so's
([deft:nineify] / [deft:util]).
Replace Document::wordAt() in plumbAtCursor() with fileTokenAt(), an
acme-style button-3 file expansion: grow over isFileChar() (alnum, '_',
and '. - + / : @', matching plan9port acme's isfilec), with ':' ending
the file name and only a digit-led ':line[:col]' suffix kept attached.
This grabs whole paths like /a/b/dpg_identity.php:584 under the caret.
Also reverts the earlier ^W change so acmeEraseWord again matches acme's
C isalnum (which excludes '_').
A corrupt or desynced stream could deliver a size[4] far larger than the
negotiated msize. recvMsg trusted it, so bodyLen=int(size)-7 became huge
(or negative when size>INT_MAX), and the following b.resize()/readExactly
corrupted the heap -> SIGBUS in the PlumbReader edit-port thread. Reject
any message whose size exceeds the negotiated msize.
Document the user-side plumber rules that make files and directories open in
Kate: the edit-port routing model, setting editor = /usr/bin/kate so files
(with their addr line/col) open in Kate instead of acme, the isdir rule that
opens a directory as a folder (plumb start kate $dir), reloading a running
plumber (9p write plumb/rules), verification, and troubleshooting. Cross-link
from PLAN.md.
Note: ~/lib/plumbing itself lives in the user's home, not the repo; this
documents how to configure it.
When a plumbed path resolves to an existing directory, onPlumbEdit now opens
it as a folder in Kate (acme-style directory open) instead of trying to open
it as a document. Kate accepts a directory on its command line and is
single-instance, so relaunching the hosting kate binary (QCoreApplication::
applicationFilePath, QStandardPaths fallback) on the dir opens it in the
running window. No in-process public API exists to open a folder as a project
(our project bridge is read-only), so this uses the documented command-line
path rather than a guessed slot. 18/18 ctest.
Kate crashed on a worker thread inside olliepalette.so near
QArrayData::deallocate. Root cause: ~Plumber did stop(); wait(2000);
delete m_reader; — but the reader can block longer than 2s (NineP connect
waitForConnected(5000), or readExactly's unbounded waitForReadyRead(-1) mid
message). When wait() timed out we deleted a still-running QThread, freeing
its QString/QByteArray members under run() → use-after-free / double free on
the reader thread.
Fixes:
- ~Plumber never deletes a running thread. stop(); wait(3000); delete only
if it finished, else hand ownership to the thread via finished->deleteLater
so it frees only after run() returns.
- NineP::readExactly waits with a bounded 2s timeout instead of forever, so a
truncated/stalled message fails cleanly and the reader stays promptly
stoppable (never force-deleted while blocked).
test_plumb_live still 4/4 (reader start/recv/clean shutdown); 18/18 ctest.
Scope X/Y to the set of documents Kate currently has open
(Application::documents()) — acme's open-window set — instead of the project
index on disk. The regex matches each buffer's local file path (or display
name for an unsaved scratch buffer). Nothing is opened or read from disk; the
live Documents are edited in place, each its own undo step.
Smaller blast radius than a project-wide sweep and closer to acme's X/Y, which
iterate open windows. Drops the ProjectIndexer dependency from the X/Y path.
Docs (SAM.md, PLAN.md) updated. 18/18 ctest green.
Kate's createToolView returns a KateMDI::ToolView that already installs its
own layout. The code created a second QVBoxLayout on it, which Qt silently
rejects, leaving SamPanel unmanaged — pinned top-left at its size hint. Add
to the existing layout when present, else create one.
Also harden SamPanel: give the program editor, Run button, and output log
minimum heights and explicit Expanding size policies so they never collapse.
Replace QRegularExpression with SamRegex (src/sam/samregex.{h,cpp}), a
faithful port of plan9port src/cmd/sam/regexp.c: a Thompson/Pike NFA.
- Leftmost-LONGEST (POSIX): matches real sam exactly, including overlapping
alternation (a|ab on 'ab' -> 'ab'), eliminating the one remaining
greedy-vs-longest divergence from PCRE.
- Linear time, no backtracking: immune to catastrophic blowup ((a*)*b over
10k 'a' returns instantly).
- sam dialect only: . * + ? | ( ) [ ] ^ $ and \ escaping (\n = newline);
^/$ per-line and ./negated-classes exclude newline, all intrinsic. No
PCRE extras (\d \w \b, lookaround, non-greedy) — sam never had them.
Port notes: shunting-yard compiler + Pike VM with per-thread capture sets and
leftmost-longest newmatch(). Fixed two porting bugs vs the C original: the
l-union aliasing of OR's left/continuation with .next, and reading the popped
subid during popator for correct capture-group ids.
SamEngine now compiles/matches via SamRegex (search, s, x/y/g/v, replacement
captures). Tests: new test_samregex (18) + updated test_samengine (31, incl.
leftmostLongestSubstitution, sam-dialect capture groups). 18/18 ctest suites.
Docs: rewrite docs/SAM.md (no more PCRE deviation; dialect + semantics are
sam's), update PLAN.md and the panel hint.
Verified against plan9port: regexp(7) defines ^ as 'beginning of a line'
and $ as 'end of a line', and sam's regexp.c BOL (p==0 || prev=='\\n') /
EOL (next char '\\n') confirm per-line anchoring. The engine was matching
with buffer-wide anchors, so ,s/^/> /g only touched the first line and the
sam idiom ,x/.+/ s/^/> / failed to prefix each line.
Fix: compile every pattern with QRegularExpression::MultilineOption. '.'
still does not cross newlines (sam-faithful; PCRE default). Add 3 regression
tests (caretIsPerLine, dollarIsPerLine, samAnchorIdiom); 30/30 engine tests.
Docs: correct docs/SAM.md — ^/$ per-line is faithful, not a deviation;
reframe PCRE extras as an escape hatch (prefer structural composition), not
'free upgrades'. Fix the panel hint and PLAN.md accordingly.
Add docs/SAM.md — a user-facing reference for the regex deviation, with
examples verified against the engine:
- leftmost-greedy (PCRE) vs leftmost-longest (sam): diverges only on
overlapping alternation (a|ab); order alternatives longest-first.
- ^/$ anchor buffer-wide, not per-line, by default. Workarounds:
structural ,x/.+/ … (preferred) or the (?m) inline flag.
- identical behaviour table (. and newlines, empty-match advance, &/\N).
- PCRE bonuses sam lacks: \d \w \b, lookahead, non-greedy, (?i)(?m)(?s).
Cross-link from PLAN.md and surface the two biggest gotchas (buffer-wide
anchors, alternation order) directly in the Sam panel hint label.
A dockable 'Sam' tool view runs the plan9 sam command language against the
active document, or project-wide via X/Y. Each Run is one undo step.
Engine (src/sam/, pure, 27 unit tests, no Kate dep):
- SamEngine::run(program,text,dotStart,dotEnd) -> edits (original-offset,
non-overlapping, sorted), dot, output, applied text. Addresses computed
against the original snapshot per sam semantics.
- Recursive-descent parser + evaluator mirroring plan9port src/cmd/sam
(cmd.c table, address.c lineaddr/charaddr, xec.c s_cmd/looper).
- Language: #n n 0 $ . ' /re/ ?re? and compound + - , ; ; commands
a c i d, s (sN, g, &, \1..\9), p =, m t, k, loops x y g v (nestable),
{ } groups, shell < > | ! via sh -c.
- peelFileLoop(): recognises a leading X/Y file loop (sam allows one per
command) for the Kate layer to drive.
Kate layer:
- SamPanel (src/plugin/sampanel): program editor + Run + output log.
- OllieView creates the tool view (MainWindow::createToolView) and owns
runSamProgram/applySamToDocument/runSamFileLoop. Single-doc edits apply
inside one Document::EditingTransaction (atomic undo). X/Y enumerates the
project index (ProjectIndexer::cachedFiles), opens each matching file, and
applies the inner program; undo is per-file (no global multi-file undo).
Regexes use QRegularExpression (PCRE), not plan9 regexp(7); documented. Out of
scope: b B n D menu, e r w f file I/O, "re" file-addressing, sam u (Kate undo).
17/17 ctest green.
Native C++ 9P2000 client over QLocalSocket (no libplumb/lib9pclient
link) talking to a running plan9port plumber:
- ninep: Tversion/Tattach/Twalk/Topen/Tread/Twrite/Tclunk, LE wire,
msize negotiated from 8192; cancellable split read (beginRead +
recvReadReply) so the edit-port reader thread honours a stop flag
without a second outstanding Tread or any cross-thread socket touch.
- plumbmsg: PlumbMsg pack/unpack with libplumb attribute quoting;
unit-tested against golden bytes captured from a live plumber.
- plumbresolve: plan9 addr (N / N:C / N.C, 1-based) -> 0-based cursor.
- plumber: send() via the send port; PlumbReader QThread on the edit
port emits edit(file,addr,wdir) via a queued signal. Namespace socket
is $NAMESPACE/plumb or /tmp/ns.$USER.$DISPLAY/plumb (plan9 getns).
Plugin: F2 (ollie_plumb) plumbs selection/wordAt(cursor) through the
send port, falling back to internal URL/file resolution; incoming edit
messages open via openUrl + activateView + setCursorPosition.
Tests: test_plumbmsg + test_plumbresolve headless; test_plumb_live does
a real send->rules->edit round-trip against a running plumber and
skips+passes when none is reachable. 16/16 ctest green; live 4/4.
Add view_close_current_space / view_close_others to isPaneFocusAction so
closing a pane warps the pointer onto the neighbour Kate activates. Same
deferred warp as split/focus.
Replace the ▯▯/⊟ glyphs with 'Split |' (left/right) and 'Split —'
(top/bottom), which render cleanly in any font. Extend isPaneFocusAction
to view_split_vert/horiz so the cursor warps onto the newly created pane
(Kate activates it), and defer the warp ~a frame (16ms) so a freshly laid-
out split pane has geometry before QCursor/mapTo reads its centre.
QCursor::setPos is a no-op on Wayland, so cursor-follows-focus did
nothing there. Add WaylandCursorWarp (src/plugin/waylandcursorwarp.{h,
cpp}): on Wayland, lock the pointer to the top-level window's wl_surface,
set_cursor_position_hint to the target pane centre (surface-local), commit,
then release — the compositor warps to the hint on unlock (the mechanism
plan9port/acme use). Panes are children of one window = one surface, so
it is a warp within that surface. Handles come from QWaylandApplication +
QPlatformNativeInterface 'surface'. X11 keeps QCursor::setPos.
Build: pointer-constraints client glue generated via ECM and compiled as
a C static lib (kcoreaddons_add_plugin silently drops .c sources); enable
C language in project(); link Qt6::GuiPrivate + wayland-client. All
guarded + a stub when the Wayland bits are absent. 13 ctest suites pass.
When a radial slice fires a go_*_split_view action (Focus Left/Right/Up/
Down/Next/Prev), warp the OS pointer onto the center of the newly active
view so the mouse follows the keyboard focus. Deferred via a 0ms timer so
Kate finishes activating the target view first. Scoped to the six pane-
focus actions only — other radial activations and ordinary view changes
(clicks, file opens) do not move the pointer.
The wedge label rect is fixed-size, so long labels crowded/overflowed.
Edit radial: Comment->Cmnt, Uncomment->Uncmnt, Uppercase->UPPER,
Lowercase->lower, Join Lines->Join, Multi-Cursor->Cursors, Go to File…->
File…, Go to Symbol…->Sym…, Command Palette…->M-x. Window radial:
Split Vertical/Horizontal->Split ▯▯ / Split ⊟, Close Split->Close,
Close Others->Only, Toggle Orientation->Rotate, focus dirs->← → ↑ ↓,
Previous->Prev. Labels only; action ids unchanged. Example config synced.
Second built-in radial 'window' bound to MiddleButton: Split Vertical/
Horizontal, Close Split, Close Others, Toggle Orientation, and a Focus
sub-branch (go_left/right/up/down/next/prev_split_view). Leaf ids are
Kate view-manager action objectNames, resolved via findActionByName.
Route a plain middle press (no left held) to the radial dispatch instead
of unconditionally consuming it: the left+middle Cut chord still fires
(left held), and the X11 middle-click primary paste is still suppressed
on release in both paths. Clear m_swallowMiddleRelease when a gesture
consumes the release.
Example config + builtinHasWindowRadialOnMiddleButton test updated.
All three palettes now span the window width. positionPalette drops its
fullWidth parameter (the centered path had no remaining callers) and
always full-widths.
Add OllieView::positionPalette(palette, fullWidth) and route all three
palettes through it. The Go-to-File and Go-to-Symbol switchers now span
the window width (minus a 12px margin) — long project-relative paths and
scope-qualified symbol labels fit without truncation. The M-x command
palette keeps its compact centered width. Replaces three duplicated
center-positioning blocks.
PaletteModel::rebuild splits the bulk score-only scan into per-thread
chunks via QtConcurrent::blockingMapped when item count exceeds
kParallelThreshold (10000). The scan is a pure map over read-only data
(FuzzyRanker::score is a stateless static; m_items not mutated); chunks
are concatenated in order and finalizeVisible imposes a deterministic
sort, so the result is bit-identical to the serial path regardless of
thread timing. Incremental (type-forward) path stays serial — it only
touches <=1000 items.
First-scan at 397k symbols (mobydick, 16 cores): ~373ms -> ~61ms.
Link Qt6::Concurrent into the palette lib. New test
parallelScanIsCorrectAndDeterministic (40k items) guards correctness.
Universal Ctags: SymbolIndex::ctagsBinary() prefers ctags-universal over
the alternatives-managed ctags (Exuberant 5.9 has no Kotlin/TS parser).
Kotlin/TypeScript now indexed — symbol count on mobydick: 17k -> 397k.
At 397k, per-keystroke ranking was 0.3-0.7s. Three structural fixes:
1. FuzzyRanker: replace std::set with vector+sort+unique; add
withRanges=false fast path that skips highlight computation during
the bulk scan (ranges materialised only for displayed rows).
2. PaletteModel: cap visible results to top-1000 (partial_sort),
score-only bulk, ranges for the top-K only. Nobody scrolls 400k.
3. Min query length (2): 1-char queries match everything and produce a
useless, expensive full scan; below the threshold the palette shows
the capped full list unranked. First real ranking at 2 chars; from
there incremental narrowing makes every keystroke sub-2ms.
Measured (mobydick 397k symbols, RelWithDebInfo, type-forward):
s: 337ms -> ~0ms | se: 507ms -> 373ms (1x first scan) |
ser: 586ms -> 1.9ms | server: 198ms -> 0.8ms.
3 new tests (incrementalMatchesFullScan, visibleResultsAreCapped,
shortQueryDoesNotRank). docs/PERF.md updated.
Two fixes for the Alt+G typing lag:
1. Default CMAKE_BUILD_TYPE to RelWithDebInfo when unspecified — the
shipped plugin was -O0. On mobydick (17k symbols) the worst
type-forward keystroke drops ~340ms -> ~29ms (~12x).
2. Debounce the PaletteWidget filter with an 80ms single-shot timer so a
fast typist's burst triggers one re-rank, not one per character.
applyQueryNow() flushes the pending query on Enter so activation uses
the latest text; activate() resets the debounce state.
The broad early-query keystrokes (1-2 chars matching most of N) are the
only ones still ~O(N); debounce ensures they are paid at most once per
pause. docs/PERF.md updated with measured numbers.
Add ProjectIndexer (project_lib, QtConcurrent + QFutureWatcher, 6 tests):
runs listFiles/listSymbols on the thread pool, delivers results on the UI
thread via filesReady/symbolsReady, caches per project root, and drops
stale results via a per-root generation counter.
Switchers now open the palette immediately — cache hit is instant, miss
shows a transient 'Indexing…' placeholder — and repopulate when the
background job finishes. The ~0.5-1s ctags stall on Alt+G no longer
blocks the UI thread. Cache invalidated on KateProjectBridge::
projectChanged. Register QList<Symbol> metatype; link Qt6::Concurrent.
Item building factored into buildFileItems/buildSymbolItems (shared by
the cache and async paths).
Add a benchmark harness (src/project/bench_switcher.cpp, manual target)
that times the Go-to-File/Symbol pipeline against a real project and
writes a metrics file. Measured on mobydick (12.2k files, 17.2k symbols);
findings in docs/PERF.md.
Bottleneck: PaletteModel::setQuery re-scored all N every keystroke
(Debug: 270-600ms/keystroke; cost grew with query length). Fix: when the
query is a pure append, re-score only the currently-visible subset — the
match set is monotonic under appending, so this is exact. Later keystrokes
drop from O(N) to O(matches) (symbols 'server' 30ms -> 9ms in Release).
Guarded by test_palettemodel::incrementalMatchesFullScan.
Also documented: shipped plugin is a Debug build (~5-8x slower than
Release), and ctags/listFiles run ~0.1-1s synchronously on open (move
off-thread = open follow-up).
Add ollie_goto_file and ollie_goto_symbol to the built-in radial's 'More'
branch, next to the command-palette escape hatch (panel-opening switchers
belong there, not the in-place top ring). onRadialActivated already
resolves both ids. Example config + a builtinDefaultHasLauncherSlices
test updated.
Alt+P (and the other launcher keys) collided with Kate's own
Alt+<letter> bindings; registered as ApplicationShortcut QActions this
makes Qt report an ambiguous shortcut and fire neither. Go nuclear: drop
the QAction shortcuts entirely and intercept Alt+X/P/G in eventFilter
(accept ShortcutOverride, act on KeyPress), scoped to our window — the
same ambiguity-proof technique used for the Acme keys. QActions kept for
objectName/label; onRadialActivated now resolves all three launcher ids
directly so the radial escape-hatch still reaches them.
Kate's project plugin already opens folders and shows a file tree, and we
now build on it via KateProjectBridge, so the parallel Ollie-side model
was redundant. Remove the Alt+O 'Open Folder as Project' action, the
sticky m_projectRoot + its $XDG_CONFIG_HOME/olliepalette/project
persistence (load/save), 'Close Project', and the ollie:project:* palette
entries/handlers. currentProjectRoot() is now two-tier: bridge baseDir
-> VCS discovery. Frees Alt+O. 12 ctest suites still pass.
Levi Neely