What’s different:

targets the send button area more conservatively
uses a safe observer loop instead of blocking init
no voice activation yet
just:
mic button next to send
record
send to Speaches
insert transcript
optional auto-send

I left voice activation out on purpose because that was the highest-risk piece for hanging mobile startup.

Use this flow:

replace the broken extension in your repo with this one
reinstall it
leave built-in STT off
set:
endpoint: https://ai.sleeplessgames.co/v1/audio/transcriptions
API key: local
model: whisper-1
language: en

After this version is confirmed stable, I’ll add voice activation as a second pass without touching startup.