The remaining steady-state cost after the wedge-proofing pass was
structural: every row of an unbounded transcript participated in every
layout, and full re-renders rebuilt all of it.
CSS pair (shared scroller + the ui/static duplicate):
- The messages scroller is BLOCK flow, not a column flexbox — flex
relayouts all items when the streaming row's height changes, O(rows)
per token; block flow dirties only the tail. The old per-child
flex-shrink pin (and the min-height:auto squish hazard it suppressed)
goes with it; inter-row rhythm moves to a sibling margin.
- overflow-anchor: none — the pane owns bottom pinning, and native
anchoring kept re-selecting an anchor inside the innerHTML-replaced
live bubble every frame.
- content-visibility: auto with contain-intrinsic-size: auto estimates
on .msg (80px) and .conv-batch (200px) rows, exempting the last two
children so the live tail never toggles skip-state mid-stream. The
`auto` keyword memoizes each row's rendered size, keeping scrollHeight
and the bottom pin stable once painted.
Transcript windowing (interactive pane):
- Full re-renders paint the most recent 300 messages, cut FORWARD to a
user-turn boundary so an assistant tool_calls message is never split
from the tool results that anchor to it. Hidden content sits behind a
"Load earlier messages" pager; each click grows the window a step and
refetches, restoring the scroll anchor by scrollHeight delta (the
rAF pin re-checks the near-bottom flag at fire time, so no
suppression is needed). Rewind/edit turn math is tail-relative and
unaffected — pinned as such.
- Live appends are bounded at the idle edge: past 900 rendered rows the
oldest rows are trimmed (again to a turn boundary), only while pinned
to the bottom — a scrolled-up user is reading the rows a trim would
remove. Trimmed content stays in /history and returns through the
pager; detached agent-card entries are swept.
The perf page gains ?window= (and the runner --perf-extra) so windowing
and containment effects can be isolated.
Measured (n=3000 history + 20-turn storm; baseline -> previous branch
-> this change): full replay 1060ms -> 238ms -> 28ms windowed / 107ms
with the window grown to the full transcript; longtasks during the run
6/1080ms -> 4/495ms -> none windowed / 4/205ms unwindowed; per-turn
live-storm cost at n=3000 now equals n=300 (~220ms harness floor) even
with all 25k nodes live — the transcript-size tax is gone. Known
degraded-mode cost: the chunk path at a fully-grown window measures
~1017ms vs the 833ms floor; the shipped windowed config sits at the
floor.
Patrick Buckley
2d174cd71c