Web Shell Composer Performance

Problem

Composer input competes with several unrelated or document-wide operations:

• intent classification previously lifted every settled draft into App;
• input highlighting rescanned the complete CodeMirror document;
• slash and mention menus rebuilt unchanged state and could render unbounded category or command lists;
• the touch textarea reread computed styles after every input;
• transcript store notifications rerendered Web Shell once per streaming chunk.

These costs are small in an empty session but compound with long drafts, large command sets, and active streaming.

Design

Keep work proportional to what is visible or changing:

• intent classification receives drafts through a stable callback and only updates React when its visible suggestion changes;
• input decorations cover CodeMirror's visible ranges;
• slash completion reads only the current line and translates its result back to document coordinates;
• command and mention menus expose at most 50 immediately rendered results and reuse semantically unchanged menu state;
• inline-tag presence checks stop at the first decoration while retaining ordinary document-change mapping;
• mobile browsers use CSS intrinsic textarea sizing when available, while the fallback caches its computed height cap;
• Web Shell subscribes to transcript blocks through an animation-frame boundary, coalescing streaming notifications before they reach App or split chat panes;
• the shared streaming-state hook subscribes to its derived enum instead of rerendering consumers for block updates that keep the same state.

Search still reaches commands and providers outside the initial 50-item view. Transcript consumers receive the newest complete snapshot at most once per animation frame.

Safety

• Suggestion cancellation, stale-draft checks, and session checks remain synchronous.
• CodeMirror rebuilds decorations when the document or viewport changes.
• The mobile JavaScript auto-grow path remains as a fallback when field-sizing: content is unsupported.
• Approval extraction, goal state, transcript callbacks, and message conversion all consume the same framed transcript snapshot.

Verification

• A draft waiting for intent classification does not rerender the app shell.
• A 10,000-line composer scans only its visible line for input decorations.
• A slash refresh in a 10,000-line draft does not serialize the document again, and multiline replacement coordinates remain correct.
• Removing the final inline tag through an ordinary document change clears attachment state.
• Slash and mention menus cap their immediate result sets at 50.
• Repeating an unchanged mention refresh preserves state identity.
• Mobile input does not reread computed styles after mount.
• One hundred transcript notifications in a frame produce one React update containing the latest snapshot.
• Equivalent transcript updates do not rerender streaming-state-only consumers.