Skip to content
Merged
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
60 changes: 60 additions & 0 deletions docs/lightspeed.feature
Original file line number Diff line number Diff line change
Expand Up @@ -31,6 +31,66 @@ Feature: Lightspeed: raea as fast as anything could be, with the best quality, m
And it updates the row here (value, date, commit) and appends a line to the log at the bottom
And when an experiment doesn't pay, it records the numbers and reverts the change. A measured no is a result

Rule: 0.1.4, the performance release (set 2026-11-13: "the v0.1.4 update will be completely performance based")
# Seven areas, each with a measured baseline before anything changes. Measured on a clone of a real workspace (9
# projects, 21,713 files, a 19 GB media library), the built app headless, a Mac with 64 GB.
Scenario: PF1. App start and data loading: seconds if not milliseconds
Then the baseline, warm, headless: projects listed 0.61 to 0.67 s after the app is spawned (its host up at
0.45 to 0.5 s); a project's feed shown 136 ms after the click with 6 cards, 260 to 282 ms with 41 (about 6 ms
a card)
And open: a cold start (the first after a reboot), the window on screen, and a feed of several hundred cards

Scenario: PF2. Thumbnails: feed items, the library, a project's Source
Then a composition's still is `raea play FILE --offscreen --seek N` in a process of its own, two at a time:
1.4 s, 1.9 s and 5.0 s for three real compositions (a ray-traced set is the 5 s one)
And fixed (raea #141): beside a tsconfig.json, which raea writes into every new project, that play threw before
its first frame, so every still there was a failed play and then a render (as slow, and the app's stage and
`raea play` did not work at all in such a folder)
And open: stills drawn in one long-lived process instead of one each (each pays the engine's start), and the
library's tiles on a real library (the test copy has none)

Scenario: PF3. Video takes the space it needs, no more
Then the baseline: 38 GB of video in 1,010 files on this Mac, most at 90 to 100 Mbit/s for 1080p, because every
review version and every shot to be joined was rendered at `--quality master`
And measured on 1080p24 film footage: default 16 Mbit/s, high 30, master 87; against the master encode the
default scores VMAF 99.5 (worst frame 95.8) and high 99.7
And fixed in the guides and the skill (raea #143, raea-skills #59): a version made for notes takes no --quality
flag, a film delivers at high, master is for delivered motion graphics and type, a replaced version is deleted
And open: the app saying what renders take and offering to make old versions smaller; whether the library's
copy of a render posted from a project is a clone of it or a second file

Scenario: PF4. Renders two to three times faster
Then the baseline on a real ray-traced film at 1080p: a 61-frame shot is 5.2 to 5.8 s, of which the frames are
about 2.2 s (25 to 30 a second: 18 to 20 ms waiting on the GPU and 14 to 22 ms making the frame's rows in
JavaScript, one after the other) and the rest is the process starting; the whole 1,666-frame film is 53 s
And the two levers that follow from it: a frame's rows made while the GPU draws the one before (they are
independent: about 1.5 to 2 times on the frames), and every composition of a file rendered in one process (a
film made a shot at a time pays about 3.3 s a shot for nothing: 33 shots, 110 s)
And fixed (raea #144): a ray-traced camera's MSAA was turned on again every frame and off again by Bevy with a
warning, a line a frame in the render's log

Scenario: PF5. What agents lose time on
Then fixed in the lane check (tools/agents/check, on this Mac): the CLI surface step tests the raea it rebuilds
(it tested an older build on PATH, and stayed failing after its own rebuild), and the skills step reads the feed
app at raea-feed's main (it read whatever branch the shared checkout had out): a lane's check now passes with
no variables set
And open, each seen on 2026-11-13: a render with <Audio> that reaches no sound is refused only after every frame
is drawn (53 s on a film); `--frames N` is ignored without a word when the file has a <Composition> (the whole
film was rendered, and to a folder that was 4.9 GB of stills); `but commit` with a long message commits
nothing and says so only in a hint; main can fail its own L1 after a GitHub merge (two stale generated pages
blocked every lane for a morning); a release's Mac build is about three hours each time and ran many times

Scenario: PF6. Video playback and looking around a scene
Then video (raea #137): frames decoded a quarter second ahead and shown when due, a faster pixel conversion, a
small-drawn video drawn from a halving made off the page's thread: 4K30's worst wait between frames 76 ms to
39, 4K60 on a page drawing 120 times a second 0 of 1,214 frames passed over
And open: frames still go through the CPU; the app's player re-renders its detail view at each timeupdate; the
scene inspector's frame rate while orbiting is not measured

Scenario: PF8. Pages and assets from this Mac open at once
Then covered so far by PF1 (a feed) and PF2 (stills); open: the Source view's lists, an asset's preview, a
model's, and the library on a real library

Rule: Lightspeed's goals (set 2026-10-19, after the four multi-day levers; the user: "just focus on lightspeed work")
Scenario: Who owns what now
Then lightspeed owns performance and optimization end to end: it builds its levers itself, not only measures them
Expand Down
Loading