Voice input could never have worked. `ChatPanel` reached for
`SpeechRecognition`, which WebView2 does not provide, so the only branch
ever taken was the one reporting "当前 WebView 不支持语音输入" — and the
architecture had specified a different route all along
(`01_Architecture_Design.md:1057-1063`): press to record, upload the
audio, let AstrBot transcribe it.
That is what this does. `MediaRecorder` captures to Opus in WebM, the
audio is uploaded through the same attachment boundary as any other file,
and it is sent with no text of its own — the audio *is* the message, and
the server's speech-to-text stage turns it into one. The `record` segment
this produces comes for free from the message-composition fix earlier
today.
The slot is fixed at the start of the recording, so a recording finished
after the user switched tabs still reaches the conversation they were
speaking to (`:1065` binds the microphone to the focused instance).
Failures are mapped rather than reported raw. Refused permission, no
microphone, and a microphone held by another program need three different
actions from the user; the old code told them "check microphone
permissions" for all of them. A WebView2 build with no audio codec is
detected before the microphone is opened rather than by letting
`MediaRecorder` throw.
The microphone track is stopped on every exit path, including a failed
encode and a recorder that cannot be constructed. Leaving it live is what
keeps Windows' recording indicator on after the user has finished
speaking, and each of those paths has a test that fails without it.
One thing this cannot settle without a real build: WebView2 raises a
`PermissionRequested` event for the microphone that the host application
handles, and Tauri registers a handler only when its `clipboard` option is
enabled, which this app does not enable. So WebView2's own default decides
whether the prompt appears, is auto-granted, or is auto-denied. The
failure mapping exists so that whichever it turns out to be, the user is
told something they can act on.
Frontend tests 47 -> 59. All five mutations of the new assertions were
confirmed to fail.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>