XType — Project History

Archive of completed development sessions. This file is not loaded by Claude Code by default — it exists as a reference for when a current session needs historical context. Active work is tracked in PLAN.md.

Project milestones

Phase	What it delivered	Sessions
Phase 0	Ollama client, ContextBuffer, Debouncer foundations	1–3
Phase A	IBus prototype validated on KDE Wayland (Python)	4–8
Phase B	Fcitx5 production engine (C++17) with browser compat	9–14
Phase D	Personalization: corpus collector, style profile, dynamic prompt	15–17
Phase U	Qt6/QML settings app — full UI prototype	U1a–U1c
Phase E	Engine-UI gap closure — config loader, all controls wired end-to-end	E0–E6

Phase 0 — Environment & Foundations

[x] Session 1 — Ollama client + latency benchmark

Files: ibus-engine/engine/inference.py, scripts/benchmark_ollama.py

Objectives:

Async Ollama HTTP client with streaming support
Cancel-on-new-request (cancel token pattern)
TTFT (time-to-first-token) measurement
Benchmark script: 20 runs against qwen2.5:0.5b, print p50/p95/p99
Verify <200ms TTFT target on local machine

Commit: feat(inference): async ollama client + TTFT benchmark — p50/p95 results

[x] Session 2 — ContextBuffer + unit tests

Files: ibus-engine/engine/context_buffer.py, ibus-engine/tests/test_context_buffer.py

Objectives:

ContextBuffer class with collections.deque, max 500 chars
Methods: append_char(), backspace(), set_suggestion(), accept_next_word(), accept_all(), dismiss()
context_text property returns full buffer as string
accept_next_word() returns next word + trailing space, advances word index, appends to buffer
append_char() and backspace() auto-dismiss any active suggestion
Full pytest suite: all state transitions, edge cases (empty buffer, single word, trailing spaces)

Commit: feat(context-buffer): state machine with full pytest suite

[x] Session 3 — Debouncer + cancellation tests

Files: ibus-engine/engine/debouncer.py, ibus-engine/tests/test_debouncer.py

Objectives:

Debouncer class: timer-based, configurable delay (default 180ms)
Each new call cancels the pending timer and restarts
Thread-safe (used from GLib main thread, fires callback on same thread via GLib.idle_add)
Tests: rapid-fire calls only trigger once, delay respected, cancel works

Commit: feat(debouncer): 180ms debounce with cancellation + tests

Phase A — IBus Prototype (Python)

[x] Session 4 — IBus engine skeleton + XML registration

Files: ibus-engine/cotypist.xml, ibus-engine/engine/main.py, ibus-engine/engine/engine.py

Objectives:

cotypist.xml: IBus component descriptor (name, exec path, description)
main.py: IBus.init(), IBus.Factory, IBus.main() loop
engine.py: CotypistEngine(IBus.EngineBase) skeleton — override do_process_key_event, do_focus_in, do_focus_out, do_reset
Engine registers with factory, loads config, initialises ContextBuffer + Debouncer
Verify: ibus-daemon --replace picks up the engine, no crashes on focus

Commit: feat(ibus): engine skeleton + XML registration

[x] Session 5 — Preedit + Tab/Escape/Backspace UX

Files: ibus-engine/engine/engine.py (extend)

Objectives:

Implement full key event flow from spec:
- Printable char → pass through, append to buffer, debounce → inference
- Tab (suggestion active) → commit next word, update preedit with remainder
- Shift+Tab → commit entire suggestion, clear preedit
- Escape → dismiss, clear preedit
- Backspace (suggestion active) → dismiss only; (no suggestion) → remove last char
- Enter → dismiss, pass through
- Modifier combos → always pass through
- Key releases → always ignore
Preedit styling: IBus.AttrType.UNDERLINE + IBus.AttrType.FOREGROUND 0x888888
Wire debouncer callback to inference, inference result calls set_suggestion() + updates preedit

Commit: feat(ibus): preedit ghost text + full key handler UX

[x] Session 6 — Focus/reset/commit edge cases

Files: ibus-engine/engine/engine.py (extend)

Objectives:

do_focus_out: MUST call commit_text() on any active preedit before clearing — text is lost otherwise
do_focus_in: clear preedit, reset context buffer
do_reset: same as focus_in
Cancel any in-flight inference on focus change
Manual test checklist: switch apps mid-suggestion, click away, Ctrl+Tab

Commit: fix(ibus): focus-out commits preedit, focus-in resets state

[x] Session 7 — Config system + blocklist

Files: ibus-engine/engine/config.py, ibus-engine/tests/test_config.py

Objectives:

TOML config loader using tomllib (stdlib, Python 3.11+)
Reads from ~/.config/cotypist-linux/config.toml, falls back to defaults
Sections: [inference] (model, ollama_host, debounce_ms, min_context_chars, context_window), [behaviour] (tab_accepts_word, passthrough_terminals, blocklist_apps)
passthrough_terminals = true + blocklist_apps → engine returns False immediately for those window classes
Ollama health check on startup: GET /api/tags, log warning if unreachable
Write default config file if missing

Commit: feat(config): TOML config + app blocklist + ollama health check

[x] Session 8 — IBus end-to-end integration test

Files: docs/testing-ibus.md, ~/.config/environment.d/ibus.conf, ibus-engine/engine/main.py

Objectives:

Write ~/.config/environment.d/ibus.conf with QT_IM_MODULE=ibus, GTK_IM_MODULE=ibus, XMODIFIERS=@im=ibus so all apps pick up IBus automatically on login (no per-launch env prefixing needed)
Add rotating file logger to main.py → ~/.local/share/xtype/engine.log (always DEBUG, 1 MB / 3 rotations) for post-test analysis
Test in Kate (Qt6, text-input-v2): ghost text appears inline, Tab accepts word-by-word
Test in Zen Browser (GTK3, text-input-v3): GTK_IM_MODULE=ibus, verify preedit
Verify Alacritty is blocklisted (no preedit, no glitches)
Verify focus-out commit in Kate and Zen Browser
Document any issues in docs/testing-ibus.md

Outcome: End-to-end validated on KDE Wayland. Known limitations documented in docs/testing-ibus.md — notably that do_focus_in_id is never called by the KWin IBus bridge (app_id always empty), making blocklist-by-app-id inactive on Wayland. Alacritty passthrough works incidentally because terminals don't route printable chars through the IM.

Commit: test(ibus): end-to-end validation on KDE Wayland — Kate + Zen + Alacritty

Phase B — Fcitx5 Production (C++17)

[x] Session 9 — CMake scaffold + addon descriptor

Files: fcitx5-engine/CMakeLists.txt, fcitx5-engine/data/xtype-addon.conf.in, fcitx5-engine/data/xtype.conf

Objectives:

CMakeLists.txt: find Fcitx5, C++17, shared library target xtype-fcitx5
xtype-addon.conf.in: addon type=SharedLibrary, category=InputMethod
xtype.conf: IM registration (name, native name, icon)
Verify: cmake -B build -G Ninja && ninja -C build produces the .so

Commit: feat(fcitx5): cmake scaffold + addon descriptor

[x] Session 10 — InferenceClient C++ (libcurl)

Files: fcitx5-engine/src/inference_client.h, fcitx5-engine/src/inference_client.cpp

Objectives:

InferenceClient class: libcurl streaming HTTP POST to Ollama /api/generate
std::thread background worker, std::atomic<bool> cancel token
Streaming: parse NDJSON response, accumulate tokens, call onToken callback
Marshal result to Fcitx5 main thread via eventDispatcher().schedule()
Same options as Python: num_predict=30, temperature=0.3, top_p=0.9, stop tokens

Commit: feat(fcitx5): libcurl inference client with streaming + cancel

[x] Session 11 — ContextBuffer C++ port

Files: fcitx5-engine/src/context_buffer.h, fcitx5-engine/src/context_buffer.cpp

Objectives:

1:1 port of Python ContextBuffer using std::deque<char>
Identical API: appendChar(), backspace(), setSuggestion(), acceptNextWord(), acceptAll(), dismiss(), contextText()
Unit tests using Catch2 or doctest (add to CMake)

Commit: feat(fcitx5): context buffer C++ port + unit tests

[x] Session 12 — XTypeEngine C++ core

Files: fcitx5-engine/src/xtype.h, fcitx5-engine/src/xtype.cpp

Objectives:

XTypeEngine : public fcitx::InputMethodEngineV2
Override keyEvent(): same logic as IBus Python engine
ic->inputPanel().setClientPreedit() with TextFormatFlag::Underline
Focus-out: dismiss preedit (clear without committing) — ghost text is AI suggestion, not user input; committing on focus-out inserts unwanted text
Wire InferenceClient callbacks to preedit updates on main thread

Commit: feat(fcitx5): full engine core — key handler + preedit + focus

[x] Session 13 — KDE Plasma Wayland integration

Files: docs/testing-fcitx5.md

Objectives:

Activate: Settings → Keyboard → Virtual Keyboard → Fcitx5 (must be KWin-launched)
Disable IBus autostart to avoid conflicts
Test matrix: Kate ✅, Firefox (GTK_IM_MODULE=fcitx) ✅, Chromium (--wayland-text-input-version=3) ⚠️, Konsole 🚫, Electron (ELECTRON_OZONE_PLATFORM_HINT=wayland) ⚠️, LibreOffice ✅
Document results in docs/testing-fcitx5.md

Commit: test(fcitx5): KDE Plasma Wayland integration matrix

[x] Session 13.5 — Browser compatibility: Zen + Chromium ghost text

Files: fcitx5-engine/src/xtype.cpp, fcitx5-engine/src/config.h

Root-cause analysis (from debug.log):

Zen Browser — no ghost text: Log lines like key 'g' prog=zen ctx_len=0 appeared on every keystroke, with ctx_len never growing. Between keystrokes, rapid deactivate prog=zen / activate prog=zen pairs appeared. activate() called resetState() which called _ctx.reset(), wiping the context buffer on every key. Zen (GTK4/Firefox Wayland IM) cycles deactivate+activate per keystroke — this is normal GTK4 behaviour, not a bug. With the original code, min_context_chars=10 was never reachable.

Chromium — ghost text commits on defocus: Log entries deactivate prog=chromium hasSuggestion=1 confirmed the commitString("") guard saw the suggestion — unlike Kate, where the suggestion was already gone by deactivate(). Yet ghost text still committed. TextFormatFlag::DontCommit is a Fcitx5-internal signal; Chromium's own text-input-v3 implementation ignores it and commits the preedit via the Wayland protocol before the engine's deactivate callback can act.

Resolution:

Zen context buffer reset: changed activate() to NOT call resetState(). New resetInferenceOnly() helper cancels in-flight inference and increments _gen but leaves _ctx intact. Only wipes _ctx in deactivate() when the program changes (tracked via _lastProg member). After fix, ctx_len grows correctly in Zen between keystrokes.
Chromium focus-out commit: call clearPreedit(ic) at the TOP of deactivate(), unconditionally. This sends an empty preedit to the Wayland compositor before Fcitx5 processes the deactivation commit. Even if Chromium then commits the preedit, it commits an empty string. commitString("") retained as belt-and-suspenders for other IM clients.
Blocklist case mismatch: config.h had {"konsole", "alacritty"} but Fcitx5 reports ic->program() as "Konsole" and "Alacritty" (capitalised). Switched isBlocked() to lowercase both sides before matching.

Outcome: Ghost text works correctly in Kate, Zen Browser, and Chromium. Konsole and Alacritty are correctly suppressed by the case-insensitive blocklist.

Commit: fix(fcitx5): browser compat — Zen context reset + Chromium focus-out commit

[x] Session 14 — Model upgrade + anti-loop inference redesign

Files: fcitx5-engine/src/inference_client.h, fcitx5-engine/src/inference_client.cpp, fcitx5-engine/src/xtype.cpp

Objectives:

Switch default model from qwen2.5:0.5b to qwen2.5:1.5b for materially stronger suggestions
Replace single-turn /api/generate with chat-format /api/chat using assistant-prefill: the beginning of the user's current context is injected as the assistant's partial reply, so the model continues it rather than echoing from token 0
Stop token strategy tightened to prevent runaway multi-sentence generation

Outcome: Ghost text quality improved significantly. The assistant-prefill pattern eliminates the echo/repetition loop that appeared with the larger model on plain generate prompts. Sessions 15–17 build on this prompt structure.

Commit: feat(fcitx5): qwen2.5:1.5b + assistant-prefill chat format — anti-loop inference redesign

Phase D — Personalization Engine

[x] Session 15 — Writing corpus collector

Files: fcitx5-engine/src/corpus_collector.h, fcitx5-engine/src/corpus_collector.cpp, fcitx5-engine/src/config.h (extend)

Objectives:

CorpusCollector runs a background thread; record() is non-blocking — pushes text to a bounded deque (cap 1000), flushed every 60 s or when queue exceeds 100 entries
Sentence-level filters at record time: drops code-shaped patterns ([{};=]{2,}), all-digit/punctuation runs, fragments under min_sentence_chars
Hard-coded blocklist in corpus_collector.cpp permanently excludes password managers regardless of user config: keepassxc, 1password, bitwarden, gnome-keyring, seahorse
Corpus size cap: atomic rename rotation (corpus.txt → corpus.txt.1) when file exceeds max_corpus_mb
Only characters from appendChar feed the corpus — acceptNextWord/acceptAll output is explicitly excluded so AI suggestions never enter training data
Opt-in by default (learning.enabled = false)

Commit: feat(personalization): writing corpus collector with opt-in logging + privacy guards

[x] Session 16 — Style profile extraction

Files: fcitx5-engine/src/style_profile.h, fcitx5-engine/src/style_profile.cpp, fcitx5-engine/tests/test_style_profile.cpp

Objectives:

StyleProfile::loadFromCorpus() streams the corpus line-by-line (no full-file slurp); selects 3–5 length-varied exemplar sentences using a seeded RNG for test determinism
Privacy filters applied before exemplar selection: rejects email-like patterns, sequences of ≥5 digits, sentences containing password/secret/token/auth (case-insensitive)
serialize()/deserialize() JSON round-trip persisted at ~/.local/share/xtype/style_profile.json
Staleness check: profile regenerated only when corpus mtime > last_updated + 30 min or 200+ new lines added — avoids CPU churn during heavy typing
Full Catch2 test suite covering sampling, privacy filters, serialize/deserialize round-trip, and empty-corpus safety

Commit: feat(personalization): style profile extraction with privacy filtering + tests

[x] Session 17 — Dynamic system prompt + observability

Files: fcitx5-engine/src/prompt_builder.h, fcitx5-engine/src/prompt_builder.cpp, fcitx5-engine/src/inference_client.h/.cpp, fcitx5-engine/src/xtype.h/.cpp

Objectives:

buildSystemPrompt() assembles in layers: base instruction → user description (About the user) → style exemplars → avoid phrases
Token budget enforced: assembled prompt > 2000 chars triggers drop-longest-exemplar-first until within budget; truncation is logged as a warning
Profile refresh timer fires every 5 minutes via eventDispatcher().schedule() on the main thread — background regeneration, main-thread swap
InferenceClient::set_system_prompt() mutex-guarded for safe updates between in-flight requests
Manual test matrix verified end-to-end on KDE Wayland and documented

Commit: feat(personalization): dynamic system prompt with style examples + profile refresh Commit: feat(personalization): observability + manual test matrix for live verification

Phase U — Qt6/QML Settings App

[x] Session U1a — Engine scaffold + Qt6 window chrome

Files: fcitx5-engine/src/engine_metrics.h, fcitx5-engine/src/recent_events.h, settings/qt6-app/CMakeLists.txt, settings/qt6-app/src/main.cpp, settings/qt6-app/qml/App.qml, settings/qt6-app/qml/Theme.qml, settings/qt6-app/qml/primitives/XSidebar.qml, settings/qt6-app/qml/primitives/XTitleBar.qml

Objectives:

Engine scaffolded with EngineMetrics and RecentEvents structs that the settings app will read from JSON (written by the engine in Phase E)
Qt6/QML project set up: CMakeLists, main.cpp, Inter + JetBrains Mono variable fonts bundled
Theme QML singleton: color palette, typography scale, grain texture provider for background
Window chrome: frameless XTitleBar (draggable), XSidebar (navigation + live metrics panel), App.qml content stack

Commit: feat(engine): scaffold config + metrics + recent-events for settings UI Commit: feat(settings): qt6 project scaffold + main + cmake Commit: feat(settings): theme singleton + grain provider + bundled fonts Commit: feat(settings): window chrome + sidebar + titlebar

[x] Session U1b — Primitives + config store + engine probe

Files: settings/qt6-app/qml/primitives/ (full library), settings/qt6-app/src/config_store.h/.cpp, settings/qt6-app/src/engine_probe.h/.cpp, settings/qt6-app/src/reloader.h/.cpp, settings/qt6-app/tests/test_config_store.cpp

Objectives:

Complete QML primitive library: XToggle, XSlider, XPill, XSegmented, XRow, XCard, XSection, XBars, XButton, XInput, XKbd, XLogList, XPhraseRow, XMiniBar, XAppRow, XEngineDot
ConfigStore C++ backend: all XTypeConfig fields exposed as QML properties; TOML round-trip via toml++; atomic save (write .tmp → rename); 8 unit tests all green
EngineProbe polls metrics.json and recent_events.json on a timer and emits change signals to the sidebar
Reloader issues fcitx5-remote -r and surfaces the result to a reload banner in the UI
Fixed ReloadCenter root type and fcitx5 state detection logic

Commit: feat(settings): primitive library (toggle, slider, pill, segmented, row, bars, …) Commit: feat(settings): config store with toml round-trip + atomic save + tests Commit: feat(settings): engine probe + reloader + live sidebar metrics

[x] Session U1c — Full page suite + KDE integration

Files: settings/qt6-app/qml/pages/ (all six pages), settings/qt6-app/data/xtype-settings.desktop, UI layout fixes throughout

Objectives:

Six pages implemented from the HTML/JSX design reference: PageGeneral, PageBlockList, PagePerApp, PageModel, PagePersonalisation, PageAbout
KDE .desktop entry registers xtype-settings in application menus
Several layout issues resolved after initial port: Flickable content area had zero width (added anchors.fill), pixelSize required an explicit int cast, Gradient.Horizontal not valid in Qt6 QML

Outcome: xtype-settings binary runs standalone and displays all pages. The engine ↔ UI filesystem bridge is functional: settings written by the app are read by the engine after fcitx5-remote -r.

Commit: feat(settings): pages — general, blocklist, per-app, model, personalisation, about Commit: feat(settings): qt6/qml standalone settings app — full prototype port

Phase E — Engine-UI Gap Closure

Root cause: XTypeEngine was default-constructing XTypeConfig and never reading ~/.config/xtype/config.toml. The settings app wrote TOML correctly; the engine ignored it. Sessions E0–E6 close every gap between what the UI exposes and what the engine applies.

[x] Session E0 — Engine TOML config loader

Files: fcitx5-engine/src/config_loader.h, fcitx5-engine/src/config_loader.cpp, fcitx5-engine/src/toml/toml.hpp (vendored), fcitx5-engine/tests/test_config_loader.cpp

Objectives:

ConfigLoader::load(path) reads all XTypeConfig sections from TOML via toml++; missing keys fall back to XTypeConfig{} defaults without error
XTypeEngine::reloadConfig() override calls loader and re-applies every field
Legacy migration: tab_accepts_word = false → partial_accept = false with a logged warning
5 loader tests: missing file, malformed TOML, full round-trip, empty file, legacy migration

Outcome: Ghost text was simultaneously broken by empty stop-token strings being forwarded to Ollama. Fixed alongside E0: empty entries are now filtered from the stop-token list before payload construction. Also discovered that fcitx5-remote -r alone does not call reloadConfig(); the addon conf must declare Configurable=True for the reload signal to reach the engine.

Commit: feat(engine): load config.toml on startup and reload Commit: fix(engine+settings): ghost text broken by empty stop token and reload wiring

[x] Session E1 — Accept key wiring + phrase blocklist + threads

Files: fcitx5-engine/src/xtype.cpp, fcitx5-engine/src/phrase_blocklist.h, fcitx5-engine/tests/test_phrase_blocklist.cpp, settings/qt6-app/qml/pages/PageGeneral.qml, settings/qt6-app/qml/pages/PageModel.qml

Objectives:

Accept key made configurable: Enter and → (Right) both handled correctly in keyEvent
PhraseBlocklist: substring match, case-insensitive, reloaded on reloadConfig(); 9 unit tests including word-boundary cases
Threads slider wires num_thread into the Ollama payload when non-zero
Settings UI: removed comingSoon guards from accept-key selector and threads slider; fixed threads slider QML binding (.valid is always falsy on QML integers — replaced with ?? 0 null-coalescing)

Commit: feat(engine): wire AcceptKey::Enter and AcceptKey::Right in keyEvent Commit: feat(engine): implement phrase blocklist substring matching Commit: feat(engine): include num_thread in Ollama payload when threads is configured Commit: feat(settings): remove comingSoon from accept-key Enter/→ and threads slider

[x] Session E2 — Observability bridge + live test input

Files: fcitx5-engine/src/xtype.cpp (metrics + events writers), fcitx5-engine/src/engine_metrics.h, fcitx5-engine/src/recent_events.h, settings/qt6-app/qml/pages/PageGeneral.qml

Objectives:

Engine writes ~/.local/share/xtype/metrics.json and recent_events.json atomically (.tmp → rename) after each inference round-trip
EngineProbe in the settings app polls both files; sidebar shows live suggestions/min, average latency, corpus line count, and recent event list
PageGeneral live test input: replaced the static typewriter animation with a real XInput field that sends text to the running engine and displays its suggestion inline

Commit: feat(observability): bridge engine metrics + recent events to settings app Commit: feat(settings): replace ghost-text typewriter with live test input on General page

[x] Session E3 — User voice profile

Files: fcitx5-engine/src/xtype.cpp (buildSystemPrompt), settings/qt6-app/qml/pages/PagePersonalisation.qml

Objectives:

Engine reads user_prompt.description, user_prompt.tone, and user_prompt.avoid_phrases from config and injects them into the assembled system prompt
Description hard-capped at 500 chars in applyPrompt() (engine) and via maximumLength on the QML textarea (settings) to protect the 2000-char prompt budget
Personalisation page: description textarea, tone segmented control, avoid-phrases pill list with add/remove

Commit: feat(engine): wire user_prompt config and voice_strength into system prompt Commit: feat(settings): user voice profile UI — description, tone, avoid phrases

[x] Session E4 — Voice strength slider

Files: fcitx5-engine/src/style_profile.cpp, settings/qt6-app/qml/pages/PagePersonalisation.qml

Objectives:

voice_strength (0.0–1.0) proportionally scales how many corpus-derived style exemplars are included in the system prompt: 0 suppresses all exemplars, 1 uses the full allocation
UI slider uses onCommitted to debounce saves — no TOML write on every frame

Commit: feat(engine): proportional voice_strength exemplar control Commit: feat(settings): enable voice-strength slider

[x] Session E5 — Corpus rolling window

Files: fcitx5-engine/src/corpus_collector.cpp, settings/qt6-app/qml/pages/PagePersonalisation.qml

Objectives:

forget_after_days: entries in corpus.txt older than N days are pruned on each background flush; 0 disables pruning
pruneOldEntries() rewrites only the lines that pass the age check — no full-file load
Personalisation page: toggle + day-count input wired through ConfigStore

Commit: feat(engine): corpus rolling window for forget_after_days Commit: feat(settings): enable forget-after-days toggle

[x] Session E6 — Per-app config overrides + advanced inference settings

Files: fcitx5-engine/src/xtype.cpp, settings/qt6-app/qml/pages/PagePerApp.qml, settings/qt6-app/qml/pages/PageModel.qml, settings/qt6-app/qml/pages/PagePersonalisation.qml, settings/qt6-app/src/config_store.h/.cpp

Objectives:

Engine reads [apps.X] TOML subtables and overlays per-app suggestion length and completion mode at activate() time
Per-App page: each known app row exposes suggestion-length and mode controls
Advanced inference knobs added to ConfigStore and wired through PageModel: temperature, top_p, stop tokens
Corpus settings section added to PagePersonalisation
comingSoon guards removed from Trigger-mode and Quantisation controls (not yet implemented; controls removed entirely rather than left dimmed)

Commit: feat(engine): apply per-app config overrides in keyEvent and requestInference Commit: feat(settings): enable per-app suggestion-length and mode controls Commit: feat(settings): ConfigStore — temperature, top_p, stop_tokens, corpus knobs Commit: feat(settings): PageModel advanced inference section — temperature, top_p, stop_tokens Commit: feat(settings): PagePersonalisation corpus settings section

Pre-release fixes

isBlocked word-boundary check: prevents partial app-name matches from over-blocking (e.g. konsole no longer blocks org.kde.konsole.desktop substrings unintentionally)
Typed-buffer overflow preserves sentence boundaries: when the 500-char buffer fills, the oldest whole sentence is dropped rather than splitting mid-word
Inference log separated from fcitx5.log: engine now writes to ~/.local/share/xtype/inference.log so the settings app LogTail can show inference activity independently
User description capped at 500 chars in both engine (applyPrompt) and settings (maximumLength) to protect prompt budget

Commit: fix(engine): isBlocked word-boundary check prevents over-blocking Commit: fix(engine): typed-buffer overflow preserves sentence boundaries Commit: fix(engine): move inference log to ~/.local/share/xtype/inference.log Commit: fix(engine/settings): cap user description at 500 chars to protect prompt budget

FilesExpand file tree

HISTORY.md

Latest commit

History

HISTORY.md

File metadata and controls

XType — Project History

Project milestones

Phase 0 — Environment & Foundations

[x] Session 1 — Ollama client + latency benchmark

[x] Session 2 — ContextBuffer + unit tests

[x] Session 3 — Debouncer + cancellation tests

Phase A — IBus Prototype (Python)

[x] Session 4 — IBus engine skeleton + XML registration

[x] Session 5 — Preedit + Tab/Escape/Backspace UX

[x] Session 6 — Focus/reset/commit edge cases

[x] Session 7 — Config system + blocklist

[x] Session 8 — IBus end-to-end integration test

Phase B — Fcitx5 Production (C++17)

[x] Session 9 — CMake scaffold + addon descriptor

[x] Session 10 — InferenceClient C++ (libcurl)

[x] Session 11 — ContextBuffer C++ port

[x] Session 12 — XTypeEngine C++ core

[x] Session 13 — KDE Plasma Wayland integration

[x] Session 13.5 — Browser compatibility: Zen + Chromium ghost text

[x] Session 14 — Model upgrade + anti-loop inference redesign

Phase D — Personalization Engine

[x] Session 15 — Writing corpus collector

[x] Session 16 — Style profile extraction

[x] Session 17 — Dynamic system prompt + observability

Phase U — Qt6/QML Settings App

[x] Session U1a — Engine scaffold + Qt6 window chrome

[x] Session U1b — Primitives + config store + engine probe

[x] Session U1c — Full page suite + KDE integration

Phase E — Engine-UI Gap Closure

[x] Session E0 — Engine TOML config loader

[x] Session E1 — Accept key wiring + phrase blocklist + threads

[x] Session E2 — Observability bridge + live test input

[x] Session E3 — User voice profile

[x] Session E4 — Voice strength slider

[x] Session E5 — Corpus rolling window

[x] Session E6 — Per-app config overrides + advanced inference settings

Pre-release fixes