Research run

Enter

Sends the question in the composer; Shift + Enter adds a line.

Escape

Stops the running research from the composer.

Tab

Moves through the steps toggle, citation chips, actions, follow ups and source rows. A focused chip lights its sentence and source like hover.

EnterorSpace

On a chip, brings its source into view and flashes it; on a source row, opens its summary and link.

Arrow UporPage UporHome

Inside the thread, releases the pin to the newest line, like scrolling up.

• The block is a labelled region; each turn is an article named by its question, and the sources rail is a labelled complementary region.
• The answer is marked aria-busy while it streams; a polite live region announces once when it is ready or stopped, with the source count, instead of every word.
• Citation chips are buttons named with the source number and title. Reading progress is exposed as text on each source number, such as Reading, 60 percent.
• The steps toggle and source rows report aria-expanded; collapsed panels are inert.
• Copy, Stop and Rerun keep their width while their labels change, and state never relies on color alone.

• Streamed text is buffered and drained on a requestAnimationFrame loop at a rate that follows the measured arrival rate, scaled by how full the buffer is and eased so bursts bend the pace instead of kinking it. Reveal snaps to whole words.
• Each commit becomes its own chunk span that develops in once, opacity with a 1.5px blur that clears, and never renders again, so earlier lines never re-split or replay. Lists, headings and chips appear already formatted; a half received citation or marker waits for its first word.
• Before the first word a quiet shimmer reads Writing the answer in the exact place the first line lands. A small dot appears at the end only when the stream stalls, and leaves with the last word.
• The thread eases toward the newest line while you are at the bottom. Scrolling up releases it and a Jump to latest pill springs in above the composer.
• Sub-queries stagger in, their spinners turn into checks on the snappy spring, and source marks pop into each query row. Source rows open with a height spring; the research steps collapse on the smooth spring as the answer starts.
• The action row is reserved from the start: Stop and a live word count cross fade into Copy, Rerun and the time and token readout, so nothing below moves.
• Reduced motion reveals whole word chunks without fades or blur, stops shimmer and spinners, and snaps panels and scrolling.

• From 880px container width the sources sit in their own scrolling rail beside the thread. Below that they fold into a strip of marks under the question that opens into the list, so the streaming text stays the last thing on screen.
• Below 520px sub-query rows drop their counts and marks, the question steps down a size, and gutters tighten.
• Layout follows the block's container, not the viewport.

• Only the chunk being revealed renders; earlier chunks are memoized, so long answers stay cheap. Hover state reaches chips and sentences through one small style rule instead of re-rendering the answer.
• The pacing loop runs only while buffered text remains and commits to React only when the word boundary moves.
• Source marks are CSS masks over small SVGs, tinted by the text color.

• Choose this for search or deep research answers where people need to see what was read and check claims against sources.
• Emit sources before the text that cites them, numbered in arrival order, and send `[n]` right after the sentence it supports; the block attaches each chip to that sentence.
• Keep the answer to the supported markdown subset: paragraphs, ### headings, - and 1. lists, **bold**, `code`, fenced code and [n] citations.
• Stop keeps exactly the words already on screen and drops the buffered rest; rerun replaces the last turn in place.
• Pass a logo URL per source for a monochrome mark; brand marks in the preview are sample content from Simple Icons, not endorsements.

The full library index for assistants is at /llms.txt.