Extensions gain composer.attribute to tag prompt edits, and the prompt draft gains a text watcher

What

The prompt draft is the text you are typing in the input box before you send it. It gains a way to record where each edit came from, and extensions (add-ons that use Claude Code's host API) can now reach it:

• The draft store takes a watchText option. When set, it is given a way to read the current text and cursor, and it is told about every change as changed({before, after, origin}).
• A new attribute(origin, fn) method tags every edit made inside fn with that origin.
• The plugin host binding gains attributeDraft, alongside the existing readDraft, fillDraft and suggestDraft.
• The host API for extensions gains composer.attribute, which calls attributeDraft, and view.show.
• The text input gets new virtualTexts and commitEdit props. Clicking and selecting text now account for virtual text decorations, meaning text shown in the box that is not part of what you typed.
• Undo now restores text inside an attributed edit, and typed edits go through a new path instead of the old recording path.

The finding did not show what supplies watchText.

Why

Extensions can now tag the prompt edits they make, so those edits can be told apart from your own typing. Nothing found shows a feature that uses this yet, so for now it is mainly something extension authors can build on.