feat: add voice dictation via whisper.cpp

[gapmiss]

Summary

• Hold-to-talk mic button on the context strip — records audio on the phone, transcribes locally via whisper.cpp, injects text into the terminal as stdin
• Fully local — no cloud APIs, audio never leaves the machine
• Optional feature — only activates when WHISPER_MODEL env var is set; zero impact otherwise

Changes

Server (packages/agent):

• POST /api/transcribe endpoint — JWT-authenticated, accepts multipart audio via multer, converts to WAV via ffmpeg, runs whisper-cli, returns transcribed text
• config.ts — added WHISPER_CPP_PATH and WHISPER_MODEL env vars

Client (packages/web):

• useDictation hook — MediaRecorder with webm/opus + mp4/aac iOS fallback
• Mic button in ContextStrip with hold-to-talk, recording/processing toast indicators, haptic feedback
• Wired into TerminalView to send transcribed text as stdin

Other:

• Added .ngrok-free.app to Vite allowedHosts (was missing, only had .ngrok-free.dev)
• README updated with voice dictation setup instructions and configuration

Dependencies

• multer (new server dependency for multipart upload)
• Requires ffmpeg and whisper-cpp on the host (both available via brew)

Test plan

• Tap mic button to start → "Recording..." toast appears, mic access granted
• Tap mic button to stop → "Transcribing..." toast, then text appears in terminal
• Tested on iOS Safari (PWA) and desktop Chrome
• Verified no impact when WHISPER_MODEL is not set (mic button still shows, server returns 503 gracefully)
• Typecheck passes on both packages

🤖 Generated with Claude Code

[my-claude-utils]

Hey @gapmiss, thanks for the PR and for taking the time to build this out! Voice dictation via whisper.cpp is a genuinely cool idea, and the local-only approach fits well with clsh's philosophy.

That said, I'm going to close this one. The CLI is still very early and I want to keep the core lean and focused on the terminal experience for now. Adding external system dependencies (ffmpeg, whisper-cpp) and a new API surface is more than I'm comfortable merging without community discussion first.

I've opened a feature request issue so the community can weigh in. If there's enough interest, we can figure out the right way to integrate it (maybe as a plugin system down the road).

If you want to use this yourself in the meantime, a fork would work great. Appreciate the contribution!