the first version of the ai console shipped the wrong conclusion. the research said stop building, an existing app already does most of this, so we tried it, hated it, and the whole thing made the pc crawl. back to the drawing board with the instructions bouncing in my head: build a nice normal app, think fifteen seconds longer than usual, no shortcuts.
the rebuild started from scratch because the old crate was the problem. fresh rust app with a real ui, service cards for speech to text, text to speech, and the llm, each with its own port, health probe, live logs, and a status pill that knows three states: stopped, running, or external. external means some other process already answers that port, and dictation still has to work against it. that case mattered, the old stack the machine already runs looks dead if you check the process instead of the port.
then the real bugs surfaced during testing. the global hotkey registered on the main thread, which could deadlock startup. background emits through the webview could block forever if the page hung, which would freeze dictation mid sentence, so every emit moved to a dedicated thread with a non blocking channel. health polling held the config lock. and the media feature was missing from the manifest, so music auto pause had compiled into a no op on windows and nobody had noticed. four real bugs, all found because the first version was embarrassing enough to force a proper rewrite instead of a patch.