Repository navigation
Add ChatGPT Web dictation to the titlebar - #438
Conversation
|
If I understand this correctly, this should probably be a plugin or at the very least an option with disabled (do not show) as the default. |
|
That makes sense. Some context behind this contribution: I injured my left arm and currently can’t use it, so typing with only my right hand would be quite unproductive. This led me to use dictation much more frequently than before, which motivated this feature. I’m already using a local wmux build with it, and it’s been helping me a lot. I’m not sure where the right boundary is for wmux—whether this belongs in the core application, as an optional feature, or as a plugin. Some terminal applications now include dictation, and with voice-driven development and vibe coding becoming more common, I thought this could be useful to others too. My goal is to contribute something helpful to the community. I’m happy to make it opt-in, disabled and hidden by default, or adjust the implementation in whatever way the maintainers consider appropriate for accepting the PR. |
- Add workspacePrefs.dictationEnabled (default false). Off, the titlebar has no microphone and the Voice tab never refreshes, so wmux never loads chatgpt.com. - Leave Start or stop dictation unbound by default instead of claiming Ctrl+Shift+S from terminal programs; drop the rev-2 shortcut promotion, whose Ctrl+Alt+V default never shipped. - Re-check the account on window focus only after the user opened the sign-in window, rather than on every focus of every window. - Drop the planning and spec drafts from docs/superpowers. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
Thanks @nicolasjandre, and sorry about your arm. I pushed one commit on top so this can ship as the opt-in feature you and @richardaubin discussed:
Type check, both builds and the full unit suite pass. Merging now; it ships in the next release. |
Summary
Add ChatGPT Web dictation to the window titlebar, immediately before Keep Computer Awake. Start recording, click Stop, and append the completed transcript to the original terminal without submitting it. This removes the repeated browser → dictate → copy → terminal → paste workflow for Codex, Claude Code, and other terminal applications.
Motivation and usage limits
The contributor prefers ChatGPT's transcription quality, especially for their Portuguese development workflow. This exposes that existing workflow at the terminal level without requiring an OpenAI API key or an agent-specific voice feature.
The integration operates the regular signed-in ChatGPT website's dictation controls and never submits a ChatGPT message or invokes a Codex task to transcribe. OpenAI describes dictation as editable text returned before sending a message and documents regular Chat Voice separately from Codex usage. These are not a promise of unlimited transcription: website availability, account restrictions, and service limits still apply. Sending the resulting prompt to a terminal agent has that agent's normal usage implications.
We checked the proposed comparison with Claude Code: Anthropic explicitly states that Claude Code dictation does not consume messages or tokens and does not count toward
/usage. This PR therefore makes no quota-saving claim over Claude Code's built-in dictation. Its benefit is transcription preference and a shared workflow across terminal applications.Boundaries
Validation
git diff --checkpass.WMUX_REQUIRE_BUILD=1: 6,528 passed, 93 skipped, 1 failed. The remaining failure is the existing Windows environment-dependentpty-manager.test.tsassertion that a realpwsh.exeexists; the same failure was observed before this feature.Usage and troubleshooting:
docs/chatgpt-web-dictation.md.