Follow-up on the mentioned live-transcription tools.
HushScribe ended up eating ~4GB RAM, probably keeping Whisper loaded in memory. Gophy had enough compile/runtime issues that I never got it working cleanly (even with Codex doing the heavy lifting).
Maybe I should build this thing after all. Or try other tools.
Mini Shai-Hulud / npm reminder:
You may not have run npm install yourself and still have tooling pulling packages via npx, npm exec, or MCP servers.
Most useful check I found:
$ grep -R -E "verbose title|verbose argv|http fetch|tarball|postinstall|run-script" ~/.npm/\_logs
Mine exposed repeated npm exec @browsermcp/mcp@latest calls.
Worth auditing anything dynamically executing @ latest right now.
This weekend I caught myself rebuilding existing software twice.
At first I skipped Cobbler because it seemed too complicated for what I needed, and building a simple PXE server sounded fun and easy. Then the requirements kept showing up: profiles, templates, remote power control. Soon enough, “simple PXE server” became “rebuild Cobbler.”
Same thing with a live-audio transcription app. Whisper plus pyannote (speaker diarization) sounded exciting, so I started building before I checked what existed. Two hours in, I found HushScribe, Gophy, and a few more.
Building is the fun part. But 15 minutes of research can save hours of rebuilding the wrong thing.
PSA for macOS dev/testing: be very careful letting coding agents run tccutil, especially with broad permissions.
Codex ran tccutil reset Accessibility instead of resetting permissions for a specific app. That wiped Accessibility grants across all apps, broke input/control permissions, triggered WindowServer/HID weirdness, and made keyboard/clicks appear partially dead until sleep/wake, followed by a full reboot.
Thankfully this happened inside a macOS VM and not on my host machine.