What variant of Codex are you using?
App (macOS)
Summary
Codex App already does a reasonable job turning voice dictation into editable composer text, but there is still a gap in the post-transcription step right before submission.
This proposal is not about transcription-time prompt configuration, real-time autocomplete, or general prompt rewriting. It is specifically about an optional post-transcription / pre-send layer for voice-derived drafts in the composer.
I'd like to propose two features for that stage:
- context-aware post-transcription cleanup
- separate pre-send reminders
Core principle: context-aware, not context-overwriting. Recent thread context should help clarify the user's wording, but it should not silently invent missing requirements or change the user's intent.
Related issues
This proposal overlaps with a few existing issues, but differs in scope:
What feels different here is the combination of:
- focusing specifically on voice-dictated drafts in Codex App
- operating only after transcription, before submission
- using a small recent thread window during cleanup
- keeping missing-information guidance separate from the cleaned-up draft
This proposal assumes App voice dictation already exists and only concerns the transcribed draft before send, not microphone support itself.
Proposal
1. Context-aware post-transcription cleanup
After dictation is transcribed into composer text, allow Codex App to apply an optional second-pass cleanup step that references a small recent window from the current thread.
This would help it:
- remove spoken-language disfluencies and filler phrases
- correct likely speech-recognition mistakes
- keep terminology consistent with the active discussion
- resolve vague references when the meaning is already clear from recent thread context
- turn spoken input into a clearer draft without adding new facts
2. Separate pre-send reminders
Before a voice-derived draft is sent, show a separate, non-blocking reminder if recent thread context suggests the draft may still be incomplete or potentially misleading.
Possible reminders could cover:
- missing key details
- unanswered questions from Codex
- omitted constraints
- likely recognition mistakes that may affect meaning
- potentially missing or conflicting details relative to recent thread context
This reminder should remain independent from the cleaned-up draft, so users can decide whether to revise the input before sending.
Guardrails
To keep this lightweight and predictable:
- run it once after dictation ends, not continuously during intermediate transcript updates
- use only a small recent window from the current thread
- default to using only the current thread, not broader file, tool, or project context
- preserve technical literals carefully, including filenames, paths, commands, flags, APIs, branch names, and quoted text
- prefer reminders over silent rewriting when confidence is low
- make the feature optional and user-reviewable before sending
Example
The example below intentionally includes both spoken-language disfluencies and likely speech-recognition mistakes.
Recent context:
- earlier in the thread, the issue was described as multiple chats appearing to show the same model after relaunch
- Codex previously asked for the Codex app version and macOS version
Raw voice transcript:
uh can you check that restart thing again in the mac app, the one where all the chats kinda show the same bottle after relaunch, and maybe look at the compaction morning first before the version stuff
Possible cleaned-up draft:
Investigate the macOS app restart issue where multiple chats appear to show the same model after relaunch, and check the compaction warning first.
Possible pre-send reminder:
You may still want to include the Codex app version and macOS version that Codex asked for earlier.
In this example:
- spoken filler such as "uh" and "kinda" is removed
- likely recognition mistakes such as "bottle" -> "model" and "morning" -> "warning" are corrected
- recent thread context is used only to clarify already-established references
- missing version details stay in the reminder instead of being silently added into the cleaned-up draft
Why this would be valuable
This could improve voice-first workflows in three ways:
- better transcript quality, by cleaning up spoken phrasing and likely recognition errors
- better draft quality, by turning raw speech into a clearer request for Codex
- better interaction efficiency, by warning users about missing context before a turn is sent
From a product perspective, this could help Codex offer a more distinctive and helpful pre-send input experience.
What variant of Codex are you using?
App (macOS)
Summary
Codex App already does a reasonable job turning voice dictation into editable composer text, but there is still a gap in the post-transcription step right before submission.
This proposal is not about transcription-time prompt configuration, real-time autocomplete, or general prompt rewriting. It is specifically about an optional post-transcription / pre-send layer for voice-derived drafts in the composer.
I'd like to propose two features for that stage:
Core principle: context-aware, not context-overwriting. Recent thread context should help clarify the user's wording, but it should not silently invent missing requirements or change the user's intent.
Related issues
This proposal overlaps with a few existing issues, but differs in scope:
Closed as not planned) is closely related on pre-send prompt rewriting.Closed as not planned) is also closely related, but focuses on real-time autocomplete / ghost-text suggestions.Closed) is moderately related on the voice-transcription side, especially around additional transcription instructions.Open) is tangentially related because it asks for task-aware composer suggestions in Codex CLI.What feels different here is the combination of:
This proposal assumes App voice dictation already exists and only concerns the transcribed draft before send, not microphone support itself.
Proposal
1. Context-aware post-transcription cleanup
After dictation is transcribed into composer text, allow Codex App to apply an optional second-pass cleanup step that references a small recent window from the current thread.
This would help it:
2. Separate pre-send reminders
Before a voice-derived draft is sent, show a separate, non-blocking reminder if recent thread context suggests the draft may still be incomplete or potentially misleading.
Possible reminders could cover:
This reminder should remain independent from the cleaned-up draft, so users can decide whether to revise the input before sending.
Guardrails
To keep this lightweight and predictable:
Example
The example below intentionally includes both spoken-language disfluencies and likely speech-recognition mistakes.
Recent context:
Raw voice transcript:
Possible cleaned-up draft:
Possible pre-send reminder:
In this example:
Why this would be valuable
This could improve voice-first workflows in three ways:
From a product perspective, this could help Codex offer a more distinctive and helpful pre-send input experience.