Holdkey v0.1.32
Windows 10/11 x64. Unsigned for now: SmartScreen shows "More info -> Run anyway" the first time.
Speech models (about 3 GB) download on first run. Nothing you say leaves your PC.
Changes
- opsreport: find out what he has posted instead of assuming he has not
- opsreport: a throttled watched-thread no longer withholds a count it cannot affect
- opsreport: refuse to publish a dogfood number we cannot stand behind
- app: raise the main window on the first request, and log without a launcher
- launch: probe where AppData writes actually land, instead of trusting the path
- site: make the hero download button primary, and measurable on its own
- ops: report 2026-09-08
- session: one utterance, one configuration
- opsreport: gate the fallback per typical dictation, and separate the ambition from the alarm
- llm: never post a dictation to a host the user did not choose
- opsreport: one fallback target, 10%, and say which measure it gates
- opsreport: stop double-counting Reddit replies seen in two feeds
- opsreport: the watchdog now catches a missed daily the same night
- history: exclude the style pass's own output from the voice sampler (M4-33)
- opsreport: registration script matches the live trigger and will not move it silently
- opsreport: a partial Reddit read is a blind spot, not a zero