Desktop app: annotate the in-app preview and attach the marked image to the active task
Problem
Visual review in the Codex desktop app is slower than it needs to be when a user is looking at the built-in preview pane. In a voice or live-review flow, the user can point at a part of the page, but Codex cannot see the pointer or receive a precise visual reference.
Requested capability
Add an annotation mode to the in-app preview pane:
- A visible annotation control in the preview toolbar.
- Mouse or pen drawing for circles, arrows, and freehand marks directly over the preview.
- A capture action that produces an annotated screenshot.
- One-click attachment of that image to the active Codex task, with an optional short note.
- The annotation is non-destructive: it overlays the preview and never changes the inspected webpage.
Why it matters
This closes the gap between visual review and agent execution. It lets a user communicate "this exact element" without hunting for selectors, describing layout positions in words, or leaving the Codex app.
First release scope
A compact first version could support freehand drawing, undo, clear, and attach-to-task. Later versions could add arrows, rectangles, text notes, and annotations that persist alongside the task timeline.
Acceptance criteria
- A user can enter and exit annotation mode from an in-app preview.
- Marks are drawn over the current preview without interacting with the underlying page.
- The user can undo or clear marks before capture.
- Capturing attaches a bounded-size image to the current task as visual context.
- The task can continue with the annotated image visible in its timeline.
This is particularly useful for voice-driven site and UI review, where a user can say "circle this" and immediately give the agent an unambiguous visual target.