cmux NEXT agent engine bench: WebKit modes stay WebKit, and --shots captures each engine over the window glass
Leo: "lawrence wants to do chromium... well how could we do glass through chromium?"
Scope now: the engine bench only. This PR first added a Debug toggle (agent.chromiumGlass, off) that made the Chromium agent chat see-through. That part was dropped from the branch (force-push at 04:56Z, before e397eec → d583e5d) because 8e36a5bd9ea (lawrencecchen, "cmux pages in Chromium are see-through like their WebKit pages") landed the same fork binding (cmux_tab_set_transparent_background, API 17) for every cmux page, with no toggle. No Swift, no tunable and no default change is left here.
What it does (scripts/cmux-next/agent-engine-bench.py, +27/-2)
- WebKit modes stay WebKit.
agent.chromiumis on by default since 69a51bf63be, so the bench'sownandpagebaselines were loading the chat in Chromium. They now pinagent.chromium: false;chromiummode pins it true. --shots DIR: light and dark screen captures (screencapture, one file per display) of the open chat in each mode, plusdebug.window_list, restoringdebug.appearancetosystemafterwards.- With
--shots, the scratch Ghostty config turns the window glass on (background-opacity = 0.7,background-blur = macos-glass-regular); an empty config is an opaque window, which hides what a page does over the glass. Without--shotsthe config stays empty, as before.
Bench numbers (same harness, cmux-mac-mini-cua, window glass on, 10 chats; taken before the Swift part was dropped)
| WebKit (shipping) | Chromium (opaque) | Chromium see-through | |
|---|---|---|---|
| Total memory with 1 chat | 640 MB | 944 MB | 956 MB |
| Total memory with 10 chats | 1010 MB | 1157 MB | 1156 MB |
| First chat, first content | 3112 ms (cold outlier; 318 ms in the opaque run) | 70 ms | 100 ms |
| Next chats, ready (median) | 37 ms | 42 ms | 43 ms |
| Typing to frame p50 / p95 | 4 / 7 ms | 7.3 / 10.3 ms | 9.7 / 11.3 ms |
| Typing to paint p95 | 8 ms | 15 ms | 18 ms |
| Fling: dropped frames / p95 frame, 3 runs | 28, 15, 27 / 33–40 ms | 2, 1, 2 / 17–19 ms | 3, 2, 2 / 16.8 ms |
See-through costs nothing measurable over opaque Chromium on these numbers.
No capture: the diff is a bench script on the fleet, with no app UI change. The private side-by-side is in the design references repo as cmux-next-chromium-glass-side-by-side.
Agent session: cosmic-gray
Agent session: Air-Blue
🤖 Generated with Claude Code