feat(web): a usage window shorter than a day, and an axis fine enough to draw it (AGT-4296)

[unohee]

TL;DR

The dashboard's shortest window was 24h, and a 24h total is dominated by whatever ran before the last deploy. Measured on vela the morning after the durable draft cache shipped:

window	draft-stage prompt cache
last 24h	34.1%
last 2h	66.8%

Only the second number described the code actually running. The first said the fix had not landed.

What was missing

parseUsageSince already accepted durations like 2h — the selector just never offered one. The real gap was somewhere to draw it: groupKeyOf bottomed out at day, so a one-hour window on the time axis is a single bar, which is not a series.

• by=hour (record.ts.slice(0, 13)), advertised in --help and the README alongside the other axes
• 1h / 2h / 6h in the window selector
• The axis is chosen from the window — hour up to 48h, day beyond, because thirty days by hour is 720 bars, a texture rather than a reading
• Hour buckets render on the reader's clock. Day keys deliberately do NOT convert: shifting one by the UTC offset relabels the day, and that alignment belongs to AGT-4293
• The card heading follows the axis actually drawn. Left fixed it read 일별 over labels like 9. 10. 23시 — on first load for every reader, since the default window is 24h

Two test properties worth naming

process.env.TZ is pinned in the client test file. Local-time rendering is invisible where UTC and local coincide, and CI runners are UTC — the mutant that skips the conversion entirely passed there while failing on a laptop in KST. Only the non-zero offset is load-bearing, not Seoul specifically. The suite is run both ways below.

usage.css has a tripwire asserting every var(--…) it references exists in tokens.css. Three undefined names shipped into that file during this work (--surface-2, --text-1, --text-2; the real names are --surface2, --fg-primary, --fg-secondary). An undefined custom property is not an error — the declaration is dropped and the rule renders with an inherited colour, so it looks plausible, and neither the build, nor oxlint, nor jsdom sees it.

Why this is half a change

Cross-axis filters and the drill-down panel were split out to AGT-4297. Three review rounds on the combined change all returned REVISE, with the same two classes recurring — a fix trading one defect for another, and a fix re-breakable with a green suite. Every finding in round 3 was in the drill-down UI; this half drew none across all three rounds. The commit gate treats a repeating finding as a signal the change is too large.

The split line is drawn so each PR is a vertical slice: filters ship with their consumer in AGT-4297, rather than landing as an API nobody calls.

Review

Layer 2 (independent subagent), three rounds. Layer 1 (openswarm review) is not run on my own commits per standing policy.

Mutants against this half:

mutant	tests killed
time axis always day	4
no local-time conversion	2
card heading fixed	2
hour key equal to day key	1
short windows removed	1

Gate

tsc --noEmit exit 0 · npm run build exit 0 · full suite 5984 passed / 10 skipped, and the same under TZ=UTC · oxlint 0 warnings on the changed files.

Closes AGT-4296.

🤖 Generated with Claude Code