Commit graph

4 commits

Author SHA1 Message Date
Danny Avila
b3a2ba96a9
🪟 perf: Skip Thread Rows Far From the Viewport (#18201)
* 🪟 perf: Skip Thread Rows Far From the Viewport

Opening or closing the sidebar beside a long thread laid out and painted
every row of it on every frame of the 300 ms slide, because the slide
changes the chat pane's width. In a 246-row thread that was 17 ms of main
thread per frame, 55 ms on a CPU four times slower, where the slide ran at
about a dozen frames a second with tasks over 100 ms.

A row more than a screen from the thread's viewport is now skipped by the
browser (`content-visibility: auto`), so a change that touches the whole
thread costs only the rows near the reader. Rows within a screen are left
exactly as they were: none of them gains the clipping and containment that
skipping implies. A skipped row keeps the size it was last laid out at, so
the scrollbar, the navigation rail and the mount anchor see the same
thread, and it stays reachable by find-in-page, focus and anchors.

The skip sits on the row element the rail and the mount anchor measure:
reading the box of an element inside a skipped row makes the browser lay
that row out again. Screenshot export lays every row out before it clones
the thread, since a clone has no remembered size.

The navigation rail measured every row on each frame of a width change.
It now waits for the width to hold still and measures once; a height
change, which is what a streaming reply is, is still followed frame by
frame. Alone that saved nothing, because its reads forced a layout the
frame needed anyway; with rows skipped it is a fifth of what remains.

* 🧮 fix: Keep Rows Laid Out Until Every Overlapping Capture Is Done

`revealOffscreenRows` handed each caller the thread's `watch` function as
its resume. Two captures that overlap (a second screenshot export started
before the first finished) each got one: the first to finish resumed
skipping while the second was still cloning the thread, and the second
then created another observer without disconnecting the first, which no
cleanup released. A resume called after the thread had unmounted created
an observer for a thread that was gone.

Each thread now counts its outstanding reveals. Skipping resumes when the
last one is done, a resume counts once however often it is called, and
does nothing once the thread has unmounted.
2026-11-16 08:47:14 -05:00
Danny Avila
461a4c2c52
🔇 fix: Stop Every Row's Read-Aloud Element From Loading the Page (#18168)
Each assistant row mounts a hidden `<audio autoplay>` for its read-aloud
button and passed it `src={audioRef.current?.src}`. An element with no
source reports `''`, so the row's second render wrote an empty `src`, and
its third wrote that back as the document's base URL. Both start a media
load: the first fails at once, the second fetches the app's HTML and fails
to decode it. A thread of 250 rows did that twice per row each time the
thread was opened and a run began.

The player assigns the source itself when speech starts, so the prop
carried nothing. Without it React never touches the attribute.

The live-run benchmarks ran with read-aloud off and so never mounted these
elements. They now leave it on, as the app does, and `speech.perf.spec.ts`
counts the media errors and requests the rows make.
2026-11-14 15:02:16 -05:00
Danny Avila
c3882c09ad
🧵 perf: Stop Re-rendering and Re-styling the Thread on Stream Events and Run Boundaries (#18148)
* 🎞️ perf: Run the In-Flight Label Sweep on the Compositor

The sweep on a running tool call, streaming reasoning and the elapsed timer
animated `background-position` under `background-clip: text`, which no
browser can hand to the compositor. Every frame of it ran style, paint and
commit on the main thread, and each paint walked the transcript around the
label, so the cost of a frame grew with the conversation and the sweep
stuttered whenever the main thread was busy.

`Shimmer` in `@librechat/client` draws the same sweep from two transforms
that cancel: a masked band slides across the label while a second copy of the
text inside it slides back, so the glyphs hold still and only the gap in the
mask moves. The copy is generated content, so the label is still one text
node. Every in-flight label renders through it.

`e2e/benchmarks-fade` measures a frame and each kind of stream event inside
an exported conversation, from a Chromium trace of the built client.

* 🧪 test: Let the Frame Benchmark Open a Chat in a Chosen Theme

* 🧵 perf: Keep the Composer, Nav Rail and Row Toolbars Out of Every Stream Event

A streamed delta rewrites one row, but each one also re-rendered the composer
form (and re-measured its textarea), the message nav rail (re-translating a
label per message), the sidebar's chat provider and the pending-steers row,
because each subscribed to the whole streaming message or the whole messages
array to read something a delta never changes.

- The chat contract and steering read a projection of the branch tail (where it
  sits, whether the run is paused) and fetch the message itself when a turn is
  built, through `useBranchTailGetter`. `ChatTurnsParams` carries
  `getLatestMessage` in place of `latestMessage`.
- The compact action and pending steers select what they use of the tail.
- The URL-prompt hook observes the history only while a submission waits on it.
- The nav rail's ribs are memoized apart from the rail.
- A row's toolbar keeps one element per glyph and stable handlers, so a send
  and its completion no longer re-render every button of every row.

`e2e/benchmarks-fade` gains a render-reason probe (`FADE_PERF_WHY`), a whole-run
profile, and windows for the send and its confirming frame.

* 🫥 perf: Conceal Rerun During a Run Instead of Unmounting It on Every Row

A run starting or ending withholds, then restores, the rerun control on every
row of the thread at once. Unmounting and remounting a button on each row made
the browser re-style the whole transcript both times: the stylesheet holds
`:has(:disabled)` rules and a message row is a `:has()` anchor, so a button
entering, leaving or flipping `disabled` inside a row invalidates that row's
entire subtree. In a 250-message thread that was 80 to 100 ms inside the send
keypress and again inside the final frame.

The control now stays mounted for as long as the row has a rerun to offer and
is concealed by visibility while a run is in flight: no focus, no pointer, no
place in the accessibility tree, and no layout where the toolbar starts at the
row's edge. Below `lg`, where the toolbar is centered, it leaves the flow as
before. It is deliberately not also `disabled`. The reply being written still
mounts no controls at all.

* 🎯 perf: Write :has() Rules So a Change in a Row Does Not Re-style the Row

Two selector shapes in the stylesheet left the browser unable to tell which
elements a `:has()` rule could reach, so it re-styled the whole subtree of
every `:has()` anchor above any change:

- Tailwind's `group-has-[…]`, which compiles to
  `.x:is(:where(.group):has(…) *)`. Message rows and their toolbars are
  `group`s, so a button mounting or unmounting in a row re-styled the row.
- `theme-disabled-within:[&_*]:…`, which compiles to `…:has(:disabled) *`.
  With it in the sheet, a `disabled` flip did the same.

Either one alone was enough: flipping `disabled` on one button per row of a
250-message thread took 100 to 110 ms, and mounting one took 88 ms. A run
starting or ending did one of those on every row.

The row-hover reveal, the header label crossfade and the toolbar's
hover-active hold now use `grouped-*` custom variants that emit
`:where(.group):has(…) .x`: the same elements at the same specificity, with a
subject the browser can read. A wrapper's parts take the disabled fill from
classes the component adds while it is disabled, through `theme-disabled-fill`,
instead of from `:has(:disabled) *`. The same two changes now cost 25 to 42 ms
and 27 ms.

`has-selectors.spec.ts` reads the stylesheet the app ships and fails on either
shape; `restyle.perf.spec.ts` measures the cost.

* 📱 fix: Keep a Concealed Rerun Hidden on Narrow Screens

The narrow-screen stylesheet forces every toolbar button visible from outside
any cascade layer, so it outranked the utilities that conceal a rerun while a
run is in flight: below 1024px the button stayed on screen and in the tab
order, inert. A concealed button now carries `data-concealed`, and that rule
leaves it alone.

Covered at a narrow and a wide viewport by what the reader gets: the settled
reply's rerun is hidden and cannot take focus during a run, and is back after.

* 🧮 perf: Skip the Pause Scan in Pending Steers When None Are Pending

The pending-steers row follows the conversation's messages, so it renders for
every streamed delta, and it resolved the run's pause state before checking
whether it had anything to show. That check walks every part of every message.
It now runs only with a steer to display.
2026-11-14 12:15:48 -05:00
Danny Avila
89a4b9eee5
🎞️ perf: Run the In-Flight Label Sweep on the Compositor (#18128)
* 🎞️ perf: Run the In-Flight Label Sweep on the Compositor

The sweep on a running tool call, streaming reasoning and the elapsed timer
animated `background-position` under `background-clip: text`, which no
browser can hand to the compositor. Every frame of it ran style, paint and
commit on the main thread, and each paint walked the transcript around the
label, so the cost of a frame grew with the conversation and the sweep
stuttered whenever the main thread was busy.

`Shimmer` in `@librechat/client` draws the same sweep from two transforms
that cancel: a masked band slides across the label while a second copy of the
text inside it slides back, so the glyphs hold still and only the gap in the
mask moves. The copy is generated content, so the label is still one text
node. Every in-flight label renders through it.

`e2e/benchmarks-fade` measures a frame and each kind of stream event inside
an exported conversation, from a Chromium trace of the built client.

* 🧪 test: Let the Frame Benchmark Open a Chat in a Chosen Theme
2026-11-14 10:16:52 -05:00