Changelog

Added

• Added model-aware remote ASR configuration for StepFun and Aliyun, including model-specific prompt, hotword, and realtime options.

Changed

• Kept the main window on its assigned macOS Space when opened from the status menu.

Fixed

• Fixed replacement dictionary entries not refreshing correctly after creation or update.

Added

• Added imported meeting audio processing with full transcription, speaker diarization, progress, cancellation, and completion notifications.
• Added editable meeting transcripts with waveform timelines, audio playback, and more durable meeting history.

Changed

• Improved meeting capture and history browsing with smoother long-transcript handling and clearer task status.
• Improved selected-text flows in Obsidian and Apple Intelligence model availability detection.

Fixed

• Fixed excessive CPU usage after recording by clearing stale model initialization state and stopping hidden overlay animations.

Added

• Added website and Sponsor links on the About page, and a Settings home Sponsor entry next to Changelog.
• Added meeting completion notifications for model and file jobs, including sandboxed drag-and-drop import results.
• Added an option to cap meeting history by count so older meetings can be cleaned up automatically.

Changed

• Made long meeting transcripts more responsive with a virtualized detail list.
• Reduced Final ASR stop-to-result latency while keeping the offline Final quality path.
• Kept MOSS dictionary Hotwords on Final only so short live windows no longer echo prompt terms mid-stream.

Fixed

• Fixed browser selection false positives for dictionary, note, and translation hotkeys so empty web inputs no longer claim those shortcuts, while real in-field and page selections still work on Accessibility-blind browsers like Arc.

Changed

• Redesigned the Feature overview cards with compact config badges, click-to-open feature pages, App Enhancement above Notes, and a non-scrolling overview layout.
• App Enhancement cards can show recent configured app icons when available.

Fixed

• Fixed selected-text hotkeys for browsers and Accessibility-limited editors, including recovering real mid-line selections in line-copy editors without inventing empty selections.
• Fixed Silero VAD MLX evaluation failures so local transcription startup throws a recoverable error instead of crashing.

Added

• Added Custom → Feature master toggles so translation, rewrite, notes, app enhancement, and meeting can be disabled from a two-column overview without affecting always-on transcription.
• Added imported meeting file analysis with bounded audio windows, full transcription, continuous speaker diarization, progress and cancellation, and history audio preference handling.

Changed

• Upgraded MLX Audio and related model runtimes so live ASR sessions end with structured STT output, and unreliable chunk timing is no longer treated as final timestamps.

Fixed

• Fixed selected-text hotkeys so editors like VS Code no longer treat the current line as a selection when nothing is actually selected.
• Fixed meeting microphone start when AVAudioEngine reports a sample rate that disagrees with the hardware input device.
• Fixed Create Hot Word so adding a term that already exists in the same scope reinforces it instead of failing with a SQLite duplicate-scope error.
• Fixed local model download authorization so security-scoped folder access is required and can be reauthorized from the main window sidebar.
• Fixed status-menu activation so the main window moves to the active Space when opened from the menu bar.
• Fixed app usage attribution so metrics track the final text injection destination and exclude meeting transcripts from branch counts.

Changed

• Reduced idle memory after local ASR and LLM use by reclaiming voice-activity, transcription, MLX cache, and allocator resources after model timeout, and by loading prequantized weights directly.

Fixed

• Prevented concurrent local ASR and LLM preparation from loading duplicate model instances, and ensured app shutdown waits for canceled model loads to finish.
• Ensured delayed resource reclamation still runs after meetings, temporary background work, model changes or deletion, and remote VAD-only sessions without resetting a newly started meeting's voice activity state.

Changed

• Further reduced idle memory after local ASR and LLM use by loading prequantized model weights without temporary graphs and returning released allocator pages more aggressively.

Fixed

• Prevented delayed VAD resource reclamation from resetting a newly started meeting and losing its initial voice activity state.

Changed

• Reduced memory retained after repeated Qwen3 local transcription by loading quantized weights directly without temporary full-model quantization graphs.

Fixed

• Prevented concurrent local ASR and LLM preparation from loading duplicate model instances, and ensured app shutdown waits for canceled model loads to finish.

Changed

• Reduced idle memory after local ASR and LLM use by releasing voice-activity, transcription, MLX cache, and allocator resources after the model timeout.

Fixed

• Prevented delayed memory reclamation from being skipped after meetings, temporary background work, model changes or deletion, and remote VAD-only sessions.