Voice-to-text dictation app with local (Nvidia Parakeet/Whisper) and cloud models (BYOK). Privacy-first and available cross-platform.
JavaScript aianthropiccross-platformgeminigroqlinuxmacosnvidiaopen-sourceopenaiparakeetspeech-to-texttranscribewhisperwindows
docs(changelog): add late-breaking 1.7.1 fixes
Gabriel Stein137743e46f93
diff --git a/CHANGELOG.md b/CHANGELOG.mdindex 4f6023768..b22d5ebab 100644--- a/CHANGELOG.md+++ b/CHANGELOG.md@@ -32,6 +32,7 @@ A follow-up to 1.7.0 with a stronger encryption key for stored secrets, end-to-e- **Custom prompts sync across windows without restart.** Storage events for `customPrompt.*` keys now propagate Control Panel edits to the Main Window immediately.- **Side-panel layout flip is scoped to the notes view with an active note.** It no longer leaks onto Home, Chat, or Upload when the window is narrow; the layout now also flips correctly when you drag the window narrow on an open note and reverts on widen.- **Dictation overlay can't steal focus on Linux WMs.** The floating icon is now `focusable: false`, the cross-platform equivalent of the `no_focus [instance="open-whispr"]` workaround Sway/i3/wlroots users were applying by hand. Mouse clicks on mic and cancel still work; auto-paste no longer breaks because the original text field stays focused.+- **Meeting-detected "Start Recording" no longer races on a fresh control panel.** The main process used to fire the `navigate-to-meeting-note` IPC after `did-finish-load` but before React mounted its listener, so the click was silently lost on a newly-created window. Now uses the same store-and-drain handshake the notification window already follows.### Reasoning & local LLMs@@ -39,6 +40,7 @@ A follow-up to 1.7.0 with a stronger encryption key for stored secrets, end-to-e- **Self-hosted dictation agent uses its own credentials** instead of silently falling back to the cleanup model's URL and API key.- **Qwen thinking is suppressed on local llama.cpp.** Both the note-formatting and streaming chat-agent paths now send `chat_template_kwargs.enable_thinking: false` (the llama.cpp-compatible flag) so Qwen3.x doesn't exhaust its token budget inside `<think>` and return empty content.- **Stop sending Ollama-only `think` field to Groq.** Groq strictly validates request bodies and rejects unknown fields, so the suppression helper now picks the right dialect per provider instead of attaching both.+- **`disableThinking` actually reaches every reasoning route.** `resolveReasoningRoute` previously returned a bare `{ kind: "cleanup" }` and the agent config omitted the flag, so `cleanupDisableThinking` and `dictationAgentDisableThinking` were silently dropped before reaching `processWithReasoningModel`. Now propagated through cleanup, BYOK cleanup, and streaming BYOK cleanup — Qwen finally respects the toggle on every STT path.- **Thinking-model responses parse correctly.** When a local llama-server response only includes `reasoning_content` (no `content` field — common for Qwen3 / DeepSeek-R1 in single-shot mode), the result is now read from that field instead of returning an empty string.- **Idle local LLM unloads from VRAM** after a timeout, freeing the GPU for other workloads.- **LAN / custom-cloud URL handling.** Accepts OpenAI-compatible URLs with or without a `/v1` suffix, and no longer produces `/v1/v1/chat/completions` when the stored endpoint already includes the suffix. Both LAN and custom-cloud paths now mirror the SDK's existing v1-suffix fallback.@@ -51,6 +53,7 @@ A follow-up to 1.7.0 with a stronger encryption key for stored secrets, end-to-e- **Groq Whisper prompt cap fixed.** Custom STT routing to `api.groq.com` is now detected by endpoint URL (not provider name) and the prompt budget is capped at 890 chars (down from 900) to leave margin on UTF-16 codepoint drift.- **No more boot-time BYOK auto-default override.** A one-shot migration was switching `cloudTranscriptionMode` to `byok` on every cold boot whenever any BYOK key existed, overriding subscribed users' UI selection. Both the persistence bug and the "any key" signal are gone.- **Parakeet health check no longer stalls.** Removed a `transcribing` flag that gated the watchdog interval but had no other reader — if any error path failed to clear it, the watchdog would skip every tick forever and miss a dead sidecar.+- **Transcription sync survives empty-text cloud rows.** The local `transcriptions` table enforces `text NOT NULL`, but the cloud allowed null/empty rows from earlier failed transcriptions. Pulling one of those would abort the entire sync loop with a `SqliteError`. Empty rows are now filtered on push, skipped on pull, and a defensive coalesce in `upsertTranscriptionFromCloud` guards against future bad inputs.### Streaming & cloud@@ -71,6 +74,7 @@ A follow-up to 1.7.0 with a stronger encryption key for stored secrets, end-to-e- **`macos-media-remote` works on macOS 15.4+.** Three independent regressions (main-thread deadlock, etc.) caused the binary to always return `NOT_PLAYING` on macOS 26.4.x. Music pause/resume during dictation now works again on the latest macOS.- **Pop!OS COSMIC is forced to XWayland.** COSMIC's `XDG_CURRENT_DESKTOP=COSMIC` fell outside the relaunch allowlist, so the app ran as a native Wayland client — breaking the orb's initial placement and drag. Now handled like GNOME and KDE.- **Terminal detection on GNOME Wayland** now uses AT-SPI2 instead of the X11 selection heuristic, restoring `Ctrl+Shift+V` auto-paste for native Wayland terminal emulators. Fixes #725.+- **Konsole on X11 auto-paste fixed.** Konsole intermittently reports no `WM_CLASS` via `xdotool`, so terminal detection went blind and fell through to `Ctrl+V` — which AI terminal agents like Codex and Claude Code interpret as image-paste, producing "No image found in clipboard" instead of pasting text. Detection now reads `/proc/<pid>/comm` as a complementary signal, and Konsole + X11 specifically routes through `xdotool windowactivate --sync key shift+Insert` (the XTest `Ctrl+Shift+V` path was being silently dropped by a long-standing focus/grab quirk). Closes #184; original patch and root-cause by @JGKle.- **Linux launcher symlinks** (e.g. `/usr/bin/open-whispr` → `/opt/OpenWhispr/open-whispr` from deb/rpm packages) no longer fail with "No such file or directory" — the wrapper now resolves the symlink target before sourcing.- **Windows: llama.cpp pinned to b8857** to keep `whisper-server.exe` (frozen at OpenWhispr/whisper.cpp 0.0.6) loading correctly. A llama.cpp release between b8861 and b9020 bumped ggml's ABI, leaving local Whisper users on 1.7.0 unable to transcribe; the download script now requests b8857 explicitly.- **Port availability check probes the wildcard address (`0.0.0.0` / `::`)** so sidecars don't false-positive on a free port when something is already bound on all interfaces. Resolved with a consolidated `serverUtils.isPortAvailable` with IPv6 probe. Closes #748.