Checkpoints that never touch your branch
How PaneScale snapshots a working copy byte for byte without moving HEAD, the index or any branch — and why rewinding refuses to overwrite your edits.
Shipping local voice with whisper.cpp
How PaneScale does dictation and voice commands without sending audio anywhere — model pinning, temporary recordings and commands we deliberately left out.
Long prompts are tiring to type. Speaking them is faster, and for many people more natural. But a microphone in a developer tool raises a fair question: where does the audio go? In PaneScale the answer is: nowhere.
Local transcription
Voice runs on whisper.cpp, a fast C++ implementation of OpenAI’s Whisper speech model that runs on an ordinary CPU. Two models are offered:
| Model | Size | Character |
|---|---|---|
| Whisper Base | 142 MiB | Fast, multilingual |
| Whisper Turbo | 547 MiB | More accurate, still multilingual |
Nothing is downloaded until you click. The whisper.cpp program comes from its official release; the models come from Hugging Face over HTTPS.
Pinned downloads
Every model is checked against a pinned SHA-256 hash and size before PaneScale uses it. If the file on the server ever changed — corrupted, replaced, truncated — PaneScale would refuse it instead of running an unknown model on your machine.
The recording lifecycle
- The microphone is captured in the app and converted to 16 kHz mono — what Whisper expects.
- The audio is written to a temporary WAV file.
- whisper.cpp transcribes it.
- The temporary file is deleted immediately.
Only the text remains. A recording can be up to ten minutes long, which covers even very long prompts.
Dictation lands where you are
The transcribed text is inserted where your cursor was: a terminal pane, the chat composer, a workflow brief, a task card. If the field can’t accept text, it goes to the clipboard so nothing is lost.
Commands, carefully
Voice commands act on PaneScale itself: change the layout, start an agent, queue something for tonight, ask what needs you. They are matched against a small set of phrases in English and German. If you almost said a command, PaneScale suggests the closest one instead of guessing.
What we left out
Some commands would be convenient and are deliberately missing: delete, approve, merge, push, deploy, publish. A misheard word should never be able to publish your code. Those actions stay clicks, behind approvals where it matters.
Try it
Settings → Voice & dictation: install whisper.cpp, download Whisper Base, pick your language. Then open the voice panel, press Space and say “show four panes”.