Spoken instructions with preview annotations: one reviewed message

[kvnloo]

Direction question

Would spoken feedback collected with existing preview annotations, reviewed and sent as one normal message fit T3 Code? If so, should it live under #22091 rather than become a separate feature?

The development loop I want is: inspect the running app → mark what needs changing → explain it aloud → review one draft → send. Dash is the inspiration, not a request to import its runtime or add another agent framework.

Triage: reuse what exists

Area	Existing work	Suggested direction
Fast feedback (“Turbo”)	Settings → Motion already exists. #22330 discusses controls that do not follow it. Annotation draw performance is tracked in #8059, with open PR #24573.	Use existing controls and support focused, measured fixes. No new Turbo toggle, global animation-speed change, or claim of faster model inference.
Annotation	Preview annotations already carry elements, regions, strokes, comments and a screenshot. #22604 proposes batched visual adjustments. #19126 and #25310 cover image annotation.	Reuse the preview annotation/composer path. Do not introduce a general image editor: #1429 was explicitly declined.
Live voice	iPhone dictation already exists. #22091 requests calls with a thread. Open PRs #14865 and #26393 explore different Codex live-voice designs. #25797 covers desktop dictation separately.	Do not start a third voice implementation. Keep dictation and duplex calls distinct, and align with whichever voice direction maintainers choose. These PRs are references, not evidence of approved scope.

The missing connection to consider

While annotating a preview, let spoken instructions belong to that annotation draft instead of immediately starting a separate agent turn:

1. Select elements or draw using the existing preview tools.
2. Record or collect the spoken instruction as editable text beside the annotation.
3. Review the screenshot, marks and instruction together. Edit or remove them before sending.
4. Attach them to the current composer draft without replacing existing content. Normal Send submits the combined feedback once.

Entering annotation mode must not start a microphone or send anything. If an explicitly started live call is active, collecting speech for the annotation must not also dispatch the same instruction through the call.

This is a behavior proposal, not a prescribed toolbar or implementation. A full-screen overlay, screenshot editor, background listening, CarPlay, and a new speech broker are outside this ask. Preserving annotation drafts across full reloads is also a separate product decision, as clarified on #20726.

Boundaries before implementation

• Bind capture and delivery to the original environment/thread/draft. Navigation must not redirect speech or a late transcript into another draft.
• Keep the existing distinction between transcription, which edits a draft, and sending an agent request.
• Cancellation, microphone denial and a disconnected environment must not discard existing draft content or silently submit it.
• State client/provider support explicitly. An iOS dictation path or browser-only voice demo does not establish web, packaged desktop and Android support.

I would keep performance fixes, annotation behavior and voice transport in separate contributions, with explicit direction/scope approval before feature PRs, per CONTRIBUTING.

The decision needed here is only whether this combined annotation-and-speech workflow is wanted, and where it should be tracked. “Keep it in #22091,” “dictation only,” or “not planned” would each be actionable. This is not a request to review or approve all linked PRs.

Status: source/docs and public discussion triage only. No T3 implementation, live-voice round trip or performance improvement is claimed. Drafted with GPT-6.1 Sol through the acpRegistry harness in T3 Code, on behalf of kvnloo.