Repository navigation
wall: draw the race on the picture; parse races[] and wakes[] - #1332
Conversation
The race moves from a bar panel onto the image.
- Each dark tile shows its own clock since the race began above its
progress ring and phase (tiles too small for text show the ring).
- A relit tile pulses its side's colour for 250 ms and shows its time
(first_frame_at - relaunch_all.started_at) in its centre for 3 s,
then in its corner until the race ends.
- One clock per half, centred low in the half: fcvm freezes in blue with
a check mark when its last tile returns; EC2 keeps counting with
"+N s and counting" (once its counted time passes fcvm's) and where
its slowest tile is (instance, OS, driver, app). A capacity wait is
not counted and is named: "not counting 21 s AWS spent finding a spot
GPU". On an upright phone the two clocks sit in a row under the image.
- When both halves are back, a card for 8 s: "fcvm back in 3.3 s · EC2
back in 27 s · 8.1× faster", from the coordinator's times, EC2 without
its capacity wait. Then the strip's last-race line takes over.
- The bar panel and Score.tsx are removed.
The last-race line falls back from the scoreboard to the newest race in
races[] (now parsed) to each half's newest single restart, so the strip
does not show dashes.
Asleep and waking: the recording's badge reads "◉ RECORDING · <date>"
and flips to "● LIVE" as the live image replaces it; the checklist card
becomes a lower third ("WAKING 16 GPUs 0:41.2" and the five phases as
dots), under the image on an upright phone. Asleep says "16 GPUs wake
in about 1:38 when you stay." from the median of fleet.power.wakes (now
parsed).
The spec line's fps is the renderers' reported fps, median over live
tiles, not the rate this page receives (1 to 5 fps on a tile that just
relit).
docs/screenshots regenerated from the e2e screenshot run.
Tests: unit tests for the clocks, chips, EC2 steps and gap, the
capacity-wait line, the end card and hand-over, the races[] fallback,
the asleep line and the badge; contract tests that races and wakes parse
from the coordinator's own snapshot states. Each fails against the
previous commit's core files. e2e race, schedule and asleep tests moved
to the on-picture elements.
|
Caution Review failedThe pull request is closed. Finishing Touches
Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out. Comment |
|
@codex review |
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 62bbc12a74
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…rols - Race result, frozen half clock, announcement and bar value read the coordinator's measured half time (relaunch_all.fcvm_ms / ec2_ms, now parsed; else the matching race in races[]) instead of wall time less the longest capacity wait. The coordinator leaves each tile's own wait out, so a tile back at 60 s without waiting beside one back at 50 s after a 30 s wait is 60 s; the page said 30 s. - A side tag is shortened to the half's name when that half, not the whole image, is under 320 px: one fcvm GPU against seven EC2 on a 960 px image left the full fcvm tag running off the left edge. - The top row's controls wrap under the title, right-aligned, when the two do not fit one row; at 701 to 860 px the isolation and menu buttons were off screen. Tests, each failing before the change: "ends a race with one card in the coordinator's numbers" (EC2 30 s against 60 s), "shortens a half's tag to its name" (full tag on a 120 px half), e2e "every control stays on screen from just above the phone breakpoint up" (three buttons past the right edge at 701 px).
|
@codex review |
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: 42a5359309
ℹ️ About Codex in GitHub
Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you
- Open a pull request for review
- Mark a draft as ready
- Comment "@codex review".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
…mage raceOnImage gives each clock a font size, the design's clamp(56px, 9vw, 160px) as 9% of the image width within 56 to 160 px, but no wider than its own half allows (six monospace characters plus 56 px of padding), with a 24 px floor; and a card width from it. A half narrower than the card moves the card in from the image's edge. The page draws the figure at that size. One fcvm GPU against fifteen EC2 on a 1600 px image centred a 144 px clock 50 px from the left edge, mostly off screen. Test: "sizes each half's clock to its own half and keeps the whole card on the image" fails on the previous race.ts (no size, no card width).
|
@codex review |
|
Codex Review: Didn't find any major issues. Another round soon, please! Reviewed commit: ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
If Codex has suggestions, it will comment; otherwise it will react with 👍. Codex can also answer questions or update the PR. Try commenting "@codex address that feedback". |
Stacked on: wall/full-bleed (#1331)
The race is now drawn on the picture: every dark tile counts, each half has one large clock, and each tile shows its own time as it relights. The strip under the image always has a race line, and waking is a lower third over the recording.
Contract. fleet/web only. Every number is the coordinator's measurement; the page formats it.
The race (
core/race.ts,core/view.ts,app/_wall/Overlays.tsx,app/wall.css)clamp(18px, tile.h/5, 56px)) above its progress ring and phase.first_frame_at - relaunch_all.started_at) in its centre for 3 s, then in its top right corner until the race ends.clamp(56px, 9vw, 160px), weight 300 while counting, 600 once frozen). fcvm freezes in blue with a check mark when its last tile returns; EC2 keeps counting in amber with+{gap} s and countingonce its counted time passes fcvm's and where its slowest tile is (instance → OS → driver → app, from the tiles' phases). While AWS has no spot GPU the clock does not count the wait and says so:not counting 21 s AWS spent finding a spot GPU.fcvm back in 3.3 s · EC2 back in 27 s · 8.1× faster(EC2's time without its capacity wait; the ratio only when both halves ran). Then the strip's line takes over.Score.tsxare gone. The strip says who started a running race ("Restart running", "Scheduled restart, running").The last race line, without dashes.
lastRaceLinereads the scoreboard, else the newest race inraces[](the coordinator's snapshot carries the last 10; the page did not parse it), else each half's newest single restart (Last single restarts: fcvm VM 1 back in 2.8 s · EC2 VM 1 back in 58 s), else nothing.Asleep and waking.
◉ RECORDING · 21 Oct 2026, 18:30 UTCand flips to● LIVEas the live image replaces it.WAKING 16 GPUs 0:41.2and the five wake phases as dots (each with its label and time). Asleep:Asleep. 16 GPUs wake in about 1:38 when you stay., from the median offleet.power.wakes(also newly parsed), or "The GPUs wake when someone watches." with none measured.Also
fps(median over live tiles), not the rate this page receives, which reads 1 to 5 fps for a tile that has just relit.docs/screenshotsregenerated from the e2e screenshot run.Review fixes (two later commits; the first also closes #1331's two findings)
relaunch_all.fcvm_ms/ec2_ms, now parsed, else the matching race inraces[]). The coordinator leaves each tile's own wait out, so a tile back at 60 s without waiting beside one back at 50 s after a 30 s wait is 60 s; the page said 30 s.Evidence
pnpm lint && pnpm typecheck && pnpm test: 221 unit tests pass. New: the race on the picture (clocks per half and per tile, chips fresh then in the corner, EC2's steps and gap, the capacity-wait line), the end card and its hand-over to the strip, the last-race fallback fromraces[], the asleep line from measured wakes, the badge flip; incontract.test.ts,racesandwakesparsed from the coordinator's own snapshot states (15_scoreboard,02_relaunch_generations,20_power).races[]fallback, the contract test forraces/wakes, and the race-on-picture tests fail (noraces, no clocks).pnpm build && FLEET_SCREENSHOT_DIR=$PWD/docs/screenshots pnpm test:e2e: 55 passed in 9.6 min. The run before it had one WebKit page crash (Target crashed) in the re-tile test; the WebKit subset then passed twice on its own and the whole suite passed. The restart race now checks a tile's clock, both half clocks, fcvm freezing while EC2 counts, a relit tile's time, EC2's steps, the end card's text against the coordinator's times, and the hand-over to the last-race line; the schedule test checks the capacity-wait line on EC2's clock; the asleep test checks the new badge and lower third.Downstream impact. The staging page during a race and while the fleet sleeps or wakes.