Skip to content

The ultimate harness for coding agents and beyond. All your agents. All your machines. One command center. Start with code, then follow your curiosity and build across disciplines: CAD, circuits, robots, games and music.

5.9 KB
Raw Blame

Claude Code scroll: why it changed, and what it measures now

Measured 2026-10-24 on an iPhone 14 (60 Hz) and a Galaxy S23 (120 Hz), profile builds, touched through the device's own dispatch (../perf/gestures.dart). Old is 3c024672b on feat/mobile-ios-android; new is the commit that added this file. How to run each benchmark: ../README.md, "Claude Code's fullscreen scroll". Raw results land in build/claude_scroll/ and are not committed.

  • Follow — how far the content moved against the finger; 100% is a native list.
  • Drift — rows the content kept moving after the finger and its fling stopped.
  • Stalled — frames the content stood still under a moving finger.
  • Ranges span the 2–3 rounds of one run.

What was wrong

The phone draws Claude Code's fullscreen scroll from a local mirror of its rows and asks Claude Code for more with wheel events and Page Up/Down. The old mirror:

  • moved the content 11–78% of the finger on the first drag after an agent opened, then drifted up to 6.5 rows. It did not know which rows scroll until Claude Code had answered a scroll;
  • followed the first flicks at 28–67%. It asked one page per round trip, and only once the finger reached the last row it knew;
  • stalled 76–287 frames on a flick back down to the newest row.

Frame rendering was not the cause. The Play build, with the old code, presented 107–116 fps on the S23 while the content moved (SurfaceFlinger --latency). In the benchmark, both builds build 96–99% of frames inside the 8.3 ms budget on slow reading drags.

What changed

In third_party/xterm/lib/src/ui/ (remote_scroll_mirror.dart, scroll_handler.dart, render.dart); entry 21 of third_party/xterm/README.autonomous.md.

Change Why
The next page is asked when the finger is within 80 ms of the last known row a fling no longer waits a round trip at each edge
Up to 3 Page Up/Down sent in one batch, where they land among rows already known one round trip instead of three
Which rows scroll is read off the prompt box once Claude Code has been quiet 300 ms the first drag moves 1:1
400 rows above fetched 500 ms after opening (was 120 rows after 1200 ms) the first flicks find rows already there
The blank row over the prompt box never counts as a scrolling row it is the same in every frame, so it looked like one
"Quiet" waits 1.25× the measured answer time, at most 1 s a slow link's late answer is not taken for a new frame
While the spring catches up, a finger move reaches the screen in the same frame removes a lag that grew with finger speed
A Page Up that moves nothing marks the top of the transcript flicks at the top stop on the top row
Frames placed by row-signature votes (bestPlacing) same answer as trying every placing, less work
The mirror exists from the first frame for Claude Code terminals the first drag is drawn from it too

Results

iPhone 14, fake Claude Code, ~100 ms an answer

Scenario Old New
First drag 29–44%, drift ≤ 4 rows 100%, drift 0.4
Reading drag 100% 100%
Flick up 28–49% 75–78%
Flick on into unfetched rows 38–45% 64–67%
Flick back down 57–63%, 76–124 stalled 100–101%, 0 stalled

iPhone 14, fake Claude Code, slow link (~200 ms round trip)

Scenario Old New
First drag 27–38% 100%
Reading drag 99–100%, 9–12 stalled 100%, 0 stalled
Flick up 31–35% 65–67%
Flick on into unfetched rows 29–30% 33–38%
Flick back down 101% 101%

iPhone 14, real app, a real agent over the real link

Its transcript was about 200 rows, so flicks reached its top in both builds.

Scenario Old New
First drag 11%, drift 6.5 rows 100%, drift 0.4
Flick up, before / after the same-frame finger move 50–51%, error p95 4.8–5.0 rows 101–102%, error p95 0

Galaxy S23, fake Claude Code

Scenario Old New
First drag 33–40%, drift 3.6–4.2 rows 100%, drift 0.3
Reading drag 100% 100%
Flick up 45–51% 72–82%
Flick on into unfetched rows 42–45% 57–73%
Flick back down 54–63%, 197–287 stalled 97–101%, 0–22 stalled

Galaxy S23, real app, a real agent with 408 rows above, Wi-Fi

First round / second round.

Scenario Old New
First drag 78%, drift 3.5 / 92%, drift 1.9 100%, drift 0.3 / 100%
Reading drag 100% / 101% 100% / 100%
Flick up 67% / 91% 97% / 103%
Flick on into unfetched rows 63% / 92% 99% / at the top
Flick back down 88%, 26 stalled / 97%, 44 stalled 101% / 101%

Galaxy S23, real app, same agent, 4G (LTE, ping 51–196 ms, average 123)

First round / second round.

Scenario Old New
First drag 77%, drift 3.4 / 92%, drift 2.0 100%, drift 0.3 / 100%
Reading drag 100% / 100% 100% / 100%
Flick up 52% / 91% 93% / 101%
Flick on into unfetched rows 46% / 63% 72% / 101%
Flick back down 63%, 207 stalled / 83%, 169 stalled 95%, 89 stalled / 101%, 0 stalled

Not measured: a very weak or lossy link (3G, packet loss), and whether 4G went peer to peer or through the relay.

Tests

  • test/remote_scroll_mirror_test.dart: 14 tests against a fake Claude Code. Covers bestPlacing against brute force, the fling page lead, reattaching from memory, trimming on a long read, prompt-box seeding and prefetch, multi-page batches, never blanking a row, getting back to the newest row, a slow link, the chase following the finger, and a busy program not being seeded. Each fix was reverted once to check its test fails.
  • flutter test on the 73 test files that touch the terminal: 1269 passed, 27 skipped, 0 failed. flutter analyze on the changed files: no issues.