Traditional-Chinese users got Simplified output. finalize_polished_text only ran
the deterministic OpenCC conversion for Raw mode / polish failures, leaving
successful polish modes to the LLM — which defaults to Simplified for Chinese.
The streaming-insert path bypassed the conversion entirely.
- finalize_polished_text: for non-translation dictation, always apply
apply_chinese_script_preference. It is a no-op for ChineseScriptPreference::Auto
(the default), so zero impact on users who didn't pick a script; for Simplified/
Traditional it now guarantees the output script instead of trusting the LLM.
- streaming_insert_eligible: gains the script preference and disables streaming
insert when it is non-Auto, so the one-shot path runs (streaming types live and
has no finished text to post-process).
Tests: polish_output_honors_chinese_script_preference (S2t actually converts;
Auto is a no-op) + streaming_disabled_for_non_auto_script_so_opencc_runs. 470 lib
tests pass.
Closes#643
- preserve audio on ASR failure + auto silent retry: always-archive recordings
(delete on success unless debug), route all engine failure arms through a single
fail_dictation -> recoverable transcribeFailed history; auto-retry re-feeds the
archived WAV via retranscribe_pcm; fix empty-transcript id mismatch (#613)
- hotkey: queue-chain -- a press buffered during processing auto-starts the next
recording (grace window distinguishes it from the #545 post-session cooldown)
- history: wire the manual re-transcribe button for transcribeFailed/emptyTranscript
entries with audio
Rebased onto beta 1.3.10; my capsule multi-monitor fix dropped (superseded by #714).
ensure_qa_window was un-gated, so it compiled on the Android target where WebviewWindowBuilder has no decorations/shadow/etc. Gate the desktop builder + WebviewUrl/WebviewWindowBuilder import to not(android/ios); add an Android/iOS stub returning get_webview_window (QA routes to main there; show_qa_window early-returns on Android so it's compile-only). Desktop (Win/macOS/Linux) already passed.
These 3 windows were declared in tauri.conf.json and created eagerly at startup, each holding a resident WKWebView content process even though they're rarely shown. Remove them from tauri.conf; create on first show via WebviewWindowBuilder (ensure_qa_window / ensure_less_computer_window / ensure_less_computer_glow_window) with config matching the old declarations exactly. QA drag-fix (make_qa_window_draggable_macos) was startup-only — re-applied on create. Drops ~3 resident WebKit processes (~30-50MB each) when idle.
Tradeoffs to verify on macOS: (1) first QA/Less-Computer open builds the webview before showing (slight first-use delay); (2) QA emit-race is benign (first emit = idle/empty = panel default); Less-Computer's first turn bubble could be missed on cold open (recovers next emit) — verify. Confirm QA still drags + positions; glow shows over fullscreen.
All 5 locales (en/ja/ko/zh-CN/zh-TW) were statically bundled into the index chunk that every one of the 5 webviews loads. Keep only zh-CN (fallback + source of truth) inline so init stays synchronous; lazy-load the other 4 as separate chunks on demand (initial non-zh-CN locale loads async after first paint, then changeLanguage; setLocalePreference awaits the locale before switching). Worst case = graceful zh-CN fallback, never broken keys. Cuts ~155KB raw / ~75KB gzip from the per-webview index; zh-CN users never load the other 4.
Build-verified: index 285KB + en/ja/ko/zh-TW separate chunks. Runtime-sanity: switch language in settings (en/ja/ko/zh-TW) to confirm it loads + applies.
During recording the audio level updates ~60Hz, re-rendering the whole Pill. The cancel/confirm SVG buttons and the pill metrics don't depend on level, so: React.memo(CircleButton) + useCallback(onCancel/onConfirm) lets the two buttons skip re-render per frame, and useMemo the os-derived metrics. Pure render optimization — identical output, no behavior/NSPanel change. Build-verified.
All 5 webviews (settings/capsule/qa/less-computer/glow) shared one ~942KB bundle, so every WebKit content process parsed+held the entire app (settings UI + LocalAsr 3.4k LOC + Style + Marketplace + chat panels). React.lazy each window root: FloatingShell (settings, 352KB) + QaPanel + LessComputerPanel + LessComputerGlow + Onboarding + AutoUpdateGate are now separate chunks loaded on demand. Capsule kept eager (tiny + latency-sensitive). Secondary webviews no longer load the settings UI → much lower resident memory per process.
Frontend-only, build-verified (chunks split, tsc+vite pass). Runtime-sanity: open settings / QA / Less Computer / glow once to confirm they still render (Suspense fallback is blank during first chunk load).
v1.3.7 fixed cross-desktop but not fullscreen: a plain NSWindow can't draw over ANOTHER app's fullscreen Space even with CanJoinAllSpaces+FullScreenAuxiliary+level (known Tauri limit: tauri#9556/#11488). The window must be a non-activating NSPanel. Add tauri-nspanel (macOS-only git dep), register the plugin, and convert the capsule via to_panel() at setup with NSNonactivatingPanel mask + FullScreenAuxiliary|CanJoinAllSpaces + level 25, per the plugin's own fullscreen example.
Note: Cargo.lock not regenerable here (no local cargo) — CI resolves the git dep; commit the lock update after the first local 'tauri dev'. Runtime-verify on macOS before release.
v1.3.6-5 fixed cross-desktop following but not fullscreen: FULL_SCREEN_AUXILIARY only lets the capsule JOIN the fullscreen Space; at alwaysOnTop's floating level (~3) it still renders behind the fullscreen window. setLevel(25) (above the menu bar, matching show_less_computer_glow which does overlay fullscreen) makes it render on top. Add a contract assertion to lock it in.
Runtime-verify on macOS: enter a fullscreen app, trigger dictation, capsule should now appear over it.
Add useRafThrottle; QaPanel/LessComputerPanel parse streamed markdown at frame-rate instead of per-token (was O(n^2)); LessComputer DOM-measure + IPC window-resize coalesced to one per frame. Capsule keyframes injected once into document.head instead of re-created every ~60Hz audio-level frame.
Backend perf: add [profile.release] (lto=thin/codegen-units=1/strip); tray mic watcher 1.5s->10s; global-hotkey dispatcher blocks on recv() instead of 250ms poll; coding_agent supervisor exits on non-macOS instead of 5s poll; cache+reuse reqwest client across utterances (net.rs) in OpenAI/Codex/Gemini providers (kills per-utterance cold TLS handshake).
fix(macos): capsule joins all Spaces + fullscreen via an absolute setCollectionBehavior(273) matching show_less_computer_glow; drop the racy Tauri set_visible_on_all_workspaces that deferred-overwrote the manual FullScreenAuxiliary bit.
chore: remove 12 orphaned coordinator/*.rs (never mod-declared -> not compiled); repoint macos-capsule-spaces contract test to the live coordinator.rs.
Muran Jiang