Files
pipelines/app/notes/chat-visible-row-construction.md

2.5 KiB

Chat visible-row construction

Finding and change

Every transcript delta projected the history and eagerly called _transcriptRow for every row, creating permission/question chip inputs and walking cached row content even when the row was far outside the viewport. Sliver layout was already lazy, but this construction work happened before slivers could skip it. A row-local Builder defers that construction until the sliver mounts the row. All rows, row keys, split anchors, selection, disclosures and Markdown rendering remain in the existing path. No Markdown subset, font or table layout changed.

Controlled comparison

Release web tool/chat_composer_probe_entry.dart, software Chrome at 560x1000, 300 history messages, the same 80 streamed text updates, and 120 wheel gestures. Alternating baseline/candidate/baseline/candidate, each with fresh phase pages. Baseline is 7ce4b1e6 (7287ba35 plus the preserved profiling tooling); candidate diff is the row-local Builder only. Other lanes held compiles/tests/captures for the four measured runs. Preliminary contended measurements were excluded.

Means of the two per-run statistics, milliseconds:

phase baseline build mean / p90 candidate build mean / p90 baseline / candidate raster mean
open 2.41 / 4.45 2.15 / 3.40 25.34 / 26.05
wheel scroll 2.21 / 6.55 2.25 / 6.70 21.32 / 21.30
stream 3.21 / 5.95 2.27 / 3.80 21.08 / 21.09

Streaming build mean is about 29% lower and build p90 about 36% lower here. Scroll and software-raster cost are effectively unchanged. This is a focused widget fixture, not full-shell timing or Android FPS evidence. Native hardware was unavailable. The change reduces a measured source of streaming overhead; it does not explain or claim to eliminate all perceived chat heaviness. Raw JSON and summary are in ignored app/build/perf-evidence/ and the private /tmp/chat-perf-quiet/ capture output.

Checks

Conversation + transcript suite: 644 passed, one stale narrow Cupertino label assertion failed (/tmp/memwatch/out.AEK2gT). That assertion required visible permission text, conflicting with the user's shipped icon-only header. Updated it to require no visible label while retaining current-value semantics and the permission-sheet interaction. Its complete 10-test file passed after the change (/tmp/memwatch/out.xWOzXj). Analyzer clean (flutter-analyze.uyyFoA). The existing suite covers long-history lazy mounting, paging/anchors, live updates, selection/copy, horizontal code/table scrolling and disclosure retention.