Add a toggle for spoken replies to typed input in the desktop app
Description
Feature request, not a bug. Today, choosing voice mode means speaking your input — speech-in and speech-out are a single coupled mode. The request is to decouple the two: let the user type normally while Claude's reply is read aloud. Two situations where this is the behavior you actually want: 1. The surroundings make speaking awkward but listening is fine — an open office, a shared room, a call you are half in. 2. The input contains code, file paths, or identifiers that dictation mangles. Speaking `proj_mt3wn65e7ur9yz` or `apps/web/src/app/api/bug-reports/[reportId]/route.ts` aloud is unreliable in a way that typing is not, so the input side of voice mode is the half that fails while the output side is the half that is useful. Concretely: a toggle (settings or a control next to the mic) for "read replies aloud" that is independent of the mic. Mic off + read-aloud on is the missing combination. --- Screenshot (attached as a description because Codmir bug reports cannot yet carry images — see note at the end) --- A Claude desktop app window, model footer reading "Opus 5 High". The conversation shows a user turn "Give me a JSON file, I'll use the other AI." and Claude's reply being dictated back. The composer at the bottom is in speaking state: the placeholder reads "Claude is speaking...", with a mic icon and a blue "Stop" button on the right — i.e. while the reply is being read aloud, the composer is given over to the voice session rather than accepting typed input. The "+" menu is open, listing: Add files or photos (⌘U), Add to project, Add from GitHub, Skills, Connectors, Design system, Plugins, Research, Web search (checked). The ⌘U shortcut identifies this as the macOS build. The screenshot is evidence of the coupling, not of a malfunction: nothing is broken, the two halves are simply not separable. Note for triage: the reporter has a PNG screenshot available on request. It could not be attached because the filing path has no image support — Codmir's own BugReport model carries an `attachments String[]` column, but no API route or MCP tool writes to it.
Steps to Reproduce
1. Open the Claude desktop app (macOS). 2. Type a message into the composer and send it — the reply is rendered as text only, with no option to have it read aloud. 3. Now start voice mode instead. The input switches to the microphone; the reply is spoken. 4. Try to keep the spoken reply while typing the input. There is no control that separates the two — the composer is occupied by the voice session ("Claude is speaking...", Stop button) for the duration of the reply.
Expected Behavior
A setting or control that turns on spoken replies independently of spoken input, so a user can type their message and have Claude's answer read aloud. The four combinations (type/type, type/speak, speak/type, speak/speak) should all be reachable; today only type/type and speak/speak are.
Actual Behavior
Speech output is only available inside voice mode, and entering voice mode switches the input to the microphone. While the reply is being spoken the composer shows "Claude is speaking..." with a Stop button rather than accepting typed input. There is no way to get a spoken reply to a typed message.
Environment
Product
Claude desktop app
Environment
Claude desktop app on macOS (the attach menu shows the ⌘U shortcut). Model: Opus 5, effort High. Reported by a user whose primary workstation is Linux; the desktop app is the macOS build.
Activity
No comments yet.
