From a789c6bd67fa94c1f39ef007fcfbe87f49f03826 Mon Sep 17 00:00:00 2001 From: "prompt.ac/@jeffrey" Date: Mon, 21 Sep 2026 09:28:16 -0700 Subject: [PATCH] Record native Frame timings and runtime verification --- slab/COMPUTER-USE-PERFORMANCE.md | 28 ++++++++++++++++++++++++---- 1 file changed, 24 insertions(+), 4 deletions(-) diff --git a/slab/COMPUTER-USE-PERFORMANCE.md b/slab/COMPUTER-USE-PERFORMANCE.md index f3e0c0e7dc..3cdbee388b 100644 --- a/slab/COMPUTER-USE-PERFORMANCE.md +++ b/slab/COMPUTER-USE-PERFORMANCE.md @@ -28,10 +28,25 @@ old pixels or retrying input. Remote SSH transport is unchanged. The transport comparison above uses a mock capture and measures process overhead, not native screenshot latency. -One initial live Frame call without OCR completed in 536 ms. Subsequent native -benchmarks encountered the login screen and `permission_needed`; they are -excluded. Successful native capture/OCR medians and a complete Captutor recording -remain unmeasured in this session. +Five alternating pairs of successful local captures without OCR, targeting the +same window, measured the previous Bash path against the new path: + +| Native Frame path | Total median | Transport overhead median | +| --- | ---: | ---: | +| Previous Bash path (`direct:true`) | 713 ms | 80 ms | +| In-process path | 624 ms | 26 ms | + +Native processing varied substantially (total samples ranged from 366 to 852 ms), +so the 13% total reduction is a small local sample, not a universal speedup. +Transport overhead is elapsed time minus the native `timings_ms.wall` field. +One subsequent fast-OCR capture took 465 ms: contour detection used 324.7 ms, +OCR 98.4 ms, and capture 18.2 ms. Contour detection is the largest remaining cost +in that sample; it is already suppressed during Captutor recordings. + +Other attempts encountered the login screen and `permission_needed`; they are +excluded. Neo had three successful pre-update captures (340–398 ms), but its +display became unavailable before the after measurement. Accurate-OCR medians +and a complete Captutor recording remain unmeasured in this session. Puppet's sample page list fell from 357 to 230 bytes (36%). Exact page IDs and URLs remain intact; `full:true` returns the original JSON state. These are byte @@ -39,6 +54,11 @@ counts, not tokenizer measurements. Earlier list compaction also covers fleet, prox, papers, calendar, and memory. All-day event end dates and memory ordering across years are preserved. +The JavaScript runtime from `e678ebb76` is installed on Blueberry and Neo, with +the prior pinned runtimes and launch-agent backups retained. The native app was +not rebuilt. All 32 focused local checks pass; all nine semantic-browser checks +also pass against Neo's installed runtime. + Reproduce browser timings (no user tabs or native input): ```sh -- 2.51.2